Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Lecture 4
Section A: The Sample Proportion as a Summary Statistic

Binary Data

Learning Objectives

Example 11

Upon completion of this lecture section you will be able to

Response to therapy in random sample of 1,000 HIV+ positive patients from a citywide clinical population 206 patients responded

Summarize a binary outcome across a group of individual observations via the sample proportion Explain why, with binary data, the sample proportion is the only summary statistic (besides sample size n) necessary to describe characteristics of the sample Compute the sample proportion based on the results of a study

Summary measure: sample proportion by

(pronounced p-hat!), given

http://inclass.kaggle.com/

Example 1

Example 1

Response to therapy in random sample of 1,000 HIV+ positive patients from a citywide clinical population (pronounced p-hat!), given

Response to therapy in random sample of 1,000 HIV+ positive patients from a citywide clinical population Why the hat? To distinguish , the sample estimate from the underlying true (population) proportion p. (which can only be estimated)

Summary measure: sample proportion by

may be called estimated proportion, estimated probability or estimated risk of responding to treatment

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Example 1

Example 1

Response to therapy in random sample of 1,000 HIV+ positive patients from a citywide clinical population The sample proportion is just a sample mean of 0/1 data

Response to therapy in random sample of 1,000 HIV+ positive patients from a citywide clinical population So with 206 of the 1,000 responding, we have xi=1 for 206 observations, and xi=0 for 794 observations. So

Generally, binary data values are given a value of x=1 for observations that have the outcome, and x=0 for observations that do not have the outcome.

Example 1

Example 1

Response to therapy in random sample of 1,000 HIV+ positive patients from a citywide clinical population

Response to therapy in random sample of 1,000 HIV+ positive patients from a citywide clinical population

Quantifying variability: There is a formula for the standard deviation of binary data

Percentiles? Well, if we know , we know the sample percentiles.

But this quantity is not particularly useful in understanding the distribution.

10

Example 1

Example 2: Maternal/Infant HIV Transmission2

Response to therapy in random sample of 1,000 HIV+ positive patients from a citywide clinical population

Randomized Trial: HIV positive pregnant women randomized to receive AZT or placebo

Visual Displays?

2Connor

E, et al. Reduction of Maternal-Infant Transmission of Human Immunodeficiency Virus Type 1 with Zidovudine Treatment. New England Journal of Medicine (1994). 331(18); 1173-1180 12

11

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Example 2: Maternal/Infant HIV Transmission

Example 3: Colorectal Cancer Screening3

Results

From Abstract

Of the 363 births whose HIV status was assessed (up to 18 months after birth) , 53 infants were HIV infected.

3Green

B, et al. An Automated Intervention With Stepped Increases in Support to Increase Uptake of Colorectal Cancer Screening: A Randomized Trial. Annals of Internal Medicine (2013). 158(5); 301-307 14

13

Example 3: Colorectal Cancer Screening

Example 3: Colorectal Cancer Screening

From Abstract

From Abstract

15

16

Summary

For quantifying the distribution of binary outcomes in a sample (and hence estimating the distribution in the population from which the sample was taken), the sample proportion is paramount

not only summarizes the percentage (probability, risk) of outcomes among a sample, it gives information about the variability of individual sample observations and the sample percentiles is the sample mean of sample observations that take on the value of 1 for observations with the outcome and 0 for observations without the outcome

Section B: Comparing Binary Outcomes Between Two (or More) Populations Using Sample Results: Risk Difference and Relative Risk

17

18

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Learning Objectives

Example 11

Upon completion of this lecture you will be able to:

Response to therapy in random sample of 1,000 HIV+ positive patients from a citywide clinical population 206 patients responded

Compute the risk difference and relative risk for comparing binary outcomes between two samples Interpret the risk difference and relative risk in a public health/personal health context Understand that the risk difference and relative risk will always agree in terms of direction, but can differ greatly in magnitude Understand that neither the risk difference alone, or the relative risk alone is sufficient to quantify the association of interest

Summary measure: sample proportion by

(pronounced p-hat!), given

http://inclass.kaggle.com/

19

20

Example 1

Example 1

Among the 1,000 subjects in the sample: 503 had CD4 counts < 250 at start of therapy, and 127 responded to therapy 497 had CD4 counts 250 at start of therapy, and 79 responded to therapy

How to summarize difference in response between the CD4 count groups?

2 X 2 Table Representation

Start with sample proportions:

CD4 < 250 Respond Not Respond 127 376 503

CD4 250 79 418 497 206 794 1,000

Example 1

Example 1

Summary Measure 1: the difference in proportions (also called risk difference , or attributable risk)

Summary Measure 2: the ratio proportions (also called relative risk, or risk ratio)

Interpretation(s): 9% greater (absolute) response to therapy in CD4<250 group as compared to CD4 250 group 9% greater absolute risk of response to therapy in CD4<250 group as compared to CD4 250 group

Interpretation(s): Those in the CD4<250 group have 1.56 times the chances (risk) of responding to therapy as compared to CD4 250 group 56% greater relative risk of response to therapy in CD4<250 group as compared to CD4 250 group

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Example 2: Maternal/Infant HIV Transmission2

Example 2: Maternal/Infant HIV Transmission2

Randomized Trial: HIV positive pregnant women randomized to receive AZT or placebo

Results

(at 18 mos) HIV+ HIV-

AZT 13 167 180

Placebo 40 143 183 53 310 363

2Connor

E, et al. Reduction of Maternal-Infant Transmission of Human Immunodeficiency Virus Type 1 with Zidovudine Treatment. New England Journal of Medicine (1994). 331(18); 1173-1180 25 26

Example 2: Maternal/Infant HIV Transmission2

Example 2

Results

Summary Measure 1: the difference in proportions (also called risk difference , or attributable risk)

Interpretation(s): 15% (absolute) reduction in HIV+ transmission to children born to mothers given AZT as compared to children born to mothers given placebo
(at 18 mos) HIV+ HIVAZT 13 167 180 Placebo 40 143 183 53 310 363

15% lower absolute risk of HIV+ transmission to children born to mothers given AZT

27

28

Example 2

Example 2

Summary Measure 2: the ratio proportions (also called relative risk, or risk ratio)

Risk Difference Versus Relative Risk: Substantive Interpretations Both measures use exact same information but give seemingly different results: (risk difference) 15% reduction in HIV transmission (relative risk) 68% reduction in HIV transmission Notice, both agree in terms of direction of association

Interpretation(s): Risk of mother/child HIV transmission for mothers given AZT is 0.32 times the chances (risk) of mother/child HIV transmission for mothers given placebo 68% lower relative risk of mother/child HIV transmission for mothers given AZT

29

30

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Example 2

Example 2

Risk Difference : Substantive Interpretation Can be interpreted as impact (assuming causation) at the population level For example: with this risk difference of -15% : In a population of 1,000 HIV pregnant positive women, wed expect to see 150 (15%) fewer mother/child transmissions if the 1,000 women were given AZT during pregnancy In a population of 50,000 HIV pregnant positive women, wed expect to see 7,500 (15%) fewer mother/child transmissions if the 50,000 women were given AZT during pregnancy

Relative Risk : Substantive Interpretation Can be interpreted as impact (assuming causation) at the individual level For example: with this relative risk of 0.32: The risk that a HIV+ mother who takes AZT during pregnancy transmits HIV to her child is 0.32 times her risk if she did not take AZT The risk that a HIV+ mother transmits HIV to her child is 68% lower if she takes AZT during pregnancy (as compared to if she were not taking AZT)

31

32

Example 3: Aspirin and CVD: Women3

Example 3

From Abstract

2X2 table and estimates

Ridker P, et al. A Randomized Trial of Low-Dose Aspirin in the Primary Prevention of Cardiovascular Disease in Women. New England Journal of Medicine (2005). 352(13); 1293-1304

33

34

Example 3

Example 3

Risk Difference

Relative Risk

0.2 % (absolute) reduction in (10-year) risk of CVD for women on low-dose aspirin therapy compared to women not on low dose therapy In a population of 100,000 women, we would expect to see 0.002*100,000=200 fewer cases of CVD (developing within 10 years) if the women were given low-dose aspirin therapy

10-year risk of CVD for 0.92 for women on low-dose aspirin regimen is 0.92 times the risk for women given placebo

A women can reduce her personal risk of CVD (developing within 10 years) by 8% if she takes a low dose of aspirin daily

35

36

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Risk Difference and Relative Risk

Risk Difference and Relative Risk

The risk difference and relative risk will always agree in term of the direction of estimated association If If If

It is possible to see a large effect with one measure, and a small effect with the other For example, if then decrease of 0.2% But : an absolute

: a relative decrease of 67%!

However, the two quantities can appear different in terms of magnitude

37

38

Example 4: HRT and Risk of CHD

Example 4: HRT and Risk of CHD

Marilyn Vos Savant takes on a serious question

Marilyn Vos Savant takes on a serious question

HRT and Risk of CHD

HRT and Risk of CHD

Proportion of Women Developing CHD (Incidence)

Risk Difference: Relative Risk:

HRT CHD No CHD 163 8,345 8,508

Placebo 122 7,980 8,102 285 16,325 16,610

Which value do you think was most quoted in the press?

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Example 5: Public Insurance Among Dialysis Patients4

Example 5: Public Insurance Among Dialysis Patients4

From Abstract

With more than 2 categories, common practice is to designate one of the categories as he reference group, and present comparisons of all other categories to this reference While the choice of reference group is arbitrary, in many cases it is purposely chosen to highlight the substantive emphasis For example, for this article written in a US published journal, the primary question of interest may be how the other 11 countries compared to the United States (with secondary interest in how these countries compared to each other)

Hirth R, et al. Out-Of-Pocket Spending And Medication Adherence Among Dialysis Patients In Twelve Countries. Health Affairs (2008). 27(1); 89-101

43

44

Summary

Risk difference ( ) and relative risk are two different estimates of the magnitude and direction of association for binary outcomes (between groups These two estimates are based on the exact same inputs and will always agree in terms of the direction of association, but not necessarily magnitude The risk difference helps to quantify the potential impact of a treatment or exposure for a group of individuals The relative risk helps quantify the potential impact of a treatment or exposure for an individual Neither estimate alone is sufficient to tell the whole story

Section C: Comparing Binary Outcomes Between Two (or More) Populations Using Sample Results: The Odds Ratio

45

46

Example 1:

Example 1

Response to therapy in random sample of 1,000 HIV+ positive patients from a citywide clinical population
CD4 <250 Respond Not Respond 127 376 503 CD4 250 79 418 497 206 794 1,000

Summary Measure 1: the difference in proportions (also called risk difference , or attributable risk)

Summary Measure 2: the ratio of proportions (also called relative risk, or risk ratio)

Start with sample proportions:

Summary Measure 3: the odds ratio (also called relative odds)

http://inclass.kaggle.com/ 47 48

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Odds

Risk Versus Odds

What is odds? The (estimated) odds of an event is the (estimated) probability of the event occurring, divided by the (estimated) probability of it not occurring:

As risk (

) increases, so does odds

49

50

Example 1:

Example 1:

What is odds? The (estimated) odds of an event is the (estimated) probability of the event occurring, divided by the (estimated) probability of it not occurring:

Odds Ratio: ratio of the odds of an event for two groups So for our data:

So for our data:

51

52

Example 1:

Example 2: Maternal/Infant HIV Transmission2

Odds Ratio: Interpretation The <250 CD4 count group has 1.75 times the odds of responding to therapy as the 250 CD4 count group The <250 CD4 count group has 75% greater odds of responding to therapy than the 250 CD4 count group

Randomized Trial: HIV positive pregnant women randomized to receive AZT or placebo

Odds Ratio: not a direct comparison of risks, but a comparison of a function of risks Relative Risk and odds ratio will always agree in terms of direction, but not always be the same value In this example, and

2Connor

E, et al. Reduction of Maternal-Infant Transmission of Human Immunodeficiency Virus Type 1 with Zidovudine Treatment. New England Journal of Medicine (1994). 331(18); 1173-1180 54

53

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Example 2: Maternal/Infant HIV Transmission

Example 2: Maternal/Infant HIV Transmission

Results

Results

(at 18 mos) HIV+ HIV-

AZT 13 167 180

Placebo 40 143 183 53 310 363

(at 18 mos) HIV+ HIV-

AZT 13 167 180

Placebo 40 143 183 53 310 363

55

56

Example 2

Example 2:

Summary Measure 1: the difference in proportions (also called risk difference , or attributable risk) Summary Measure 2: the ratio proportions (also called relative risk, or risk ratio)

Odds Ratio: Interpretation The AZT group has 0.27 times the odds (of HIV to child transmission) of the placebo group The AZT group has 73% lower odds of HIV to child transmission than the placebo group

Summary Measure 3: the odds ratio

Relative Risk versus Odds Ratio: in this example the relative risk and odds ratio are and

57

58

Odds Ratio: Substantive Interpretation

Relative Risk Versus Odds Ratio

Odds Ratio: Substantive Interpretation

Both measures use exact same information but can give numerically different results : both will always agree in terms of direction of association ie:

As with the relative risk, the odds ratio can be interpreted as impact (assuming causation) at the individual level The odds ratio does not directly compare the probabilities (risks, proportions) of an outcome, but instead compares a function of risk: the odds

59

60

10

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Relative Risk Versus Odds Ratio

Example 3: Aspirin and CVD: Women3

The smaller

and

are, the closer in value are

From Abstract

Ridker P, et al. A Randomized Trial of Low-Dose Aspirin in the Primary Prevention of Cardiovascular Disease in Women. New England Journal of Medicine (2005). 352(13); 1293-1304

61

62

Example 3

Example 3

2X2 table and estimates

Risk Difference

Relative Risk

Odds Ratio

63

64

Example 3:

Example 3:

Odds Ratio: Interpretation The aspirin group has 0.92 times the odds (of developing CHD) of the placebo group The aspirin group has 8% lower odds of developing CHD than the placebo group

Odds Ratio: Interpretation The aspirin group has 0.92 times the odds (of developing CHD) of the placebo group The AZT group has 8% lower odds of developing CHD than the placebo group

Relative Risk versus Odds Ratio: in this example the relative risk and odds ratio estimates are identical in value, unlike the previous two examples

Relative Risk versus Odds Ratio: in this example the relative risk and odds ratio estimates are identical in value, unlike the previous two examples

65

66

11

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Why Even Bother With The Odds Ratio?

Comparing More Than Two Groups

In many ways, the odds ratio is less intuitive and less direct measure of association than the relative risk However: In some types of studies (case control, more details coming in term 2), the odds ratio is the only measure of association that can be estimate In logistic regression (also coming in term 2), the results are initially presented as odds ratios, and hence frequently presented as odds ratios in publications

With more than 2 categories, common practice is to designate one of the categories as he reference group, and present comparisons of all other categories to this reference While the choice of reference group is arbitrary, in many cases it is purposely chosen to highlight the substantive emphasis

67

68

Example 4: Obesity and Depression

Example 4: Obesity and Depression

From Abstract

Table 4

Onyike C, et al. Is Obesity Associated with Major Depression? Results form the Third National Health and Nutrition Examination Survey. American Journal of Epidemiology (2003). 158(11); 1138-1304

69

70

Example 4: Obesity and Depression

Example 4: Obesity and Depression

Table 4

Table 4

From footnotes section:

From footnotes section:

71

72

12

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Summary

The odds ratio, ,provides an alternative to the relative risk, , for quantifying the association between a binary outcome between groups The odds ratio Is ratio of odds between two groups: odds is related to risk (probability, proportion) The odds ratio and relative risk both estimate the association between a binary outcome between groups at the individual level These two measures will agree in terms of direction, but not always magnitude The smaller the risk in the groups being compared, the more similar and

Section D: A Brief Note About Ratios

73

74

Learning Objectives

Example 1 : Maternal/Infant HIV Transmission

Upon completion of this lecture you will be able to:

Recall the Results

Understand that the scaling of ratios is not symmetric around the value of 1 (which would indicate equal values in the numerator and denominator) Consider the implications of the previous point when interpreting size of association Understand the on the log scale (well use natural log, ln) the values of ln(ratios) are symmetric about the value 0
(at 18 mos) HIV+ HIVAZT 13 167 180 Placebo 40 143 183 53 310 363

75

76

Example 1

Example 1

Summary Measures:

Interpretations: AZT associated with:

(risk difference)15% (absolute) decrease in HIV transmission risk (relative risk) 68% (relative) decrease in HIV transmission risk (odds ratio) 73% (relative) reduction in HIV transmission odds

As compared to No Treatment (Placebo)

77

78

13

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Example 1

Example 1

Direction is arbitrary: supposed we instead compared placebo to AZT:

Interpretations: No Treatment (Placebo) associated with:

(risk difference) 15% (absolute) increase in HIV transmission risk (relative risk) 210% (relative) increase in HIV transmission risk (odds ratio) 270 % (relative) increase in HIV transmission odds

As compared to AZT

79

80

Example 1

Scale of Ratios is Not Symmetric

Recap

So why do these associations seem to differ in magnitude if the direction of comparison is reversed?? The range of possible values for positive and negative associations are very different

81

82

Scale of Ratios is Not Symmetric

Scale of Ratios is Not Symmetric

The ranges are equal on the ln(Ratio) scale

For example, with relative risk: AZT to Placebo Placebo to AZT

83

84

14

Lecture 4: Statistical Reasoning for Public Health: Estimation, Inference, & Interpretation

Scale of Ratios is Not Symmetric

Summary

For example, with relative risk: AZT to Placebo Placebo to AZT

On the ratio scale (relative risk or odds ratio), the range of possible values is 0 ratio < 1: for negative associations, i.e. where the group in the numerator has lower risk (and hence odds) than the group in the denominator 1 < ratio : for positive associations, i.e. where the group in the numerator has lower risk (and hence odds) than the group in the denominator On the ln(ratio) scale, the range of possible values is: < ln(ratio) < 0: for negative associations, i.e. where the group in the numerator has lower risk (and hence odds) than the group in the denominator 0 < ln(ratio) : for positive associations, i.e. where the group in the numerator has lower risk (and hence odds) than the group in the denominator
86

85

Summary

These properties of ratios and ln(ratios) have potential implications for: Displaying associations for different group comparisons Performing statistical inference on ratios

STAY TUNED: more to come on these properties

87

15