1 of 119

Analytic Design Theory: Framework for Alignment Between Analyst and Audience

Lucy D’Agostino McGowan

Wake Forest University

2 of 119

lucymcgowan.com/talk

3 of 119

analyticdesigntheory.org

4 of 119

5 of 119

statistical thinking

6 of 119

statistical thinking

data

7 of 119

statistical thinking

data

method

8 of 119

statistical thinking

data

answer

9 of 119

https://onishlab.colostate.edu/summer-statistics-workshop-2019/which_test_flowchart/

10 of 119

design thinking

11 of 119

design thinking

Model proposed by the Hasso-Plattner Institute of Design at Stanford

12 of 119

design thinking

empathize

Model proposed by the Hasso-Plattner Institute of Design at Stanford

13 of 119

design thinking

empathize

define

Model proposed by the Hasso-Plattner Institute of Design at Stanford

14 of 119

design thinking

empathize

define

ideate

Model proposed by the Hasso-Plattner Institute of Design at Stanford

15 of 119

design thinking

empathize

define

ideate

prototype

Model proposed by the Hasso-Plattner Institute of Design at Stanford

16 of 119

design thinking

empathize

define

ideate

prototype

test

Model proposed by the Hasso-Plattner Institute of Design at Stanford

17 of 119

design thinking

empathize

define

ideate

prototype

test

Model proposed by the Hasso-Plattner Institute of Design at Stanford

18 of 119

Why?

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

19 of 119

Why?

  • Provide a common language

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

20 of 119

Why?

  • Provide a common language
  • Improve pedagogy

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

21 of 119

Why?

  • Provide a common language
  • Improve pedagogy
  • Improve alignment between data analysis producers and consumers

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

22 of 119

Producer of

Data Analysis

Consumer of

Data Analysis

Data analysis product

Data analysis evaluation

23 of 119

Producer of

Data Analysis

Consumer of

Data Analysis

Data analysis product

Data analysis evaluation

you

24 of 119

Producer of

Data Analysis

Consumer of

Data Analysis

Data analysis product

Data analysis evaluation

you

clinician

25 of 119

Producer of

Data Analysis

Consumer of

Data Analysis

Data analysis product

Data analysis evaluation

you

another statistician

26 of 119

Producer of

Data Analysis

Consumer of

Data Analysis

Data analysis product

Data analysis evaluation

you

general public

27 of 119

Producer of

Data Analysis

Consumer of

Data Analysis

Data analysis product

Data analysis evaluation

producer’s design principles

consumer’s design principles

28 of 119

Producer of

Data Analysis

Consumer of

Data Analysis

Data analysis product

Data analysis evaluation

producer’s design principles

consumer’s design principles

29 of 119

design principles

for data analysis

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

30 of 119

design principles

for data analysis

data matching

How well does the available data match the data needed to investigate a question?

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

31 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

32 of 119

33 of 119

P(Infection | Vaccination)

34 of 119

P(Vaccination | Infection)

35 of 119

36 of 119

design principles

for data analysis

data matching

How well does the available data match the data needed to investigate a question?

exhaustive

Are specific questions addressed using multiple, complementary methods, tooling or workflows?

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

37 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

38 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

39 of 119

design principles

for data analysis

data matching

How well does the available data match the data needed to investigate a question?

exhaustive

Are specific questions addressed using multiple, complementary methods, tooling or workflows?

skeptical

Are multiple, related explanations considered using the same data?

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

40 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

41 of 119

x

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

42 of 119

x

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

43 of 119

design principles

for data analysis

data matching

How well does the available data match the data needed to investigate a question?

exhaustive

Are specific questions addressed using multiple, complementary methods, tooling or workflows?

skeptical

Are multiple, related explanations considered using the same data?

second-order

Does the analysis include anything that does not directly address the primary question, but gives important context to the analysis?

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

44 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

45 of 119

design principles

for data analysis

data matching

How well does the available data match the data needed to investigate a question?

exhaustive

Are specific questions addressed using multiple, complementary methods, tooling or workflows?

skeptical

Are multiple, related explanations considered using the same data?

second-order

Does the analysis include anything that does not directly address the primary question, but gives important context to the analysis?

clarity

Does the analysis summarize data in a way that is influential in explaining how the underlying data connects to the conclusions?

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

46 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

47 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

48 of 119

design principles

for data analysis

data matching

How well does the available data match the data needed to investigate a question?

exhaustive

Are specific questions addressed using multiple, complementary methods, tooling or workflows?

skeptical

Are multiple, related explanations considered using the same data?

second-order

Does the analysis include anything that does not directly address the primary question, but gives important context to the analysis?

clarity

Does the analysis summarize data in a way that is influential in explaining how the underlying data connects to the conclusions?

reproducible

Could someone who is not the original producer take the published code and data and compute the same results?

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

49 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

50 of 119

Wake Forest University

54 Students

8 Assignments

10 point scoring

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

51 of 119

Observed between and within person variation of principles across assignment

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

52 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

53 of 119

Observed variation between principles, suggesting they measure different underlying characteristics of a data analysis

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

54 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

55 of 119

Johns Hopkins University

15 Students

2 Different Analysts

10 point scoring

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

56 of 119

The scoring of principles has some ability to distinguish between analyses done by independent analysts

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

57 of 119

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

58 of 119

Producer of

Data Analysis

Consumer of

Data Analysis

Data analysis product

Data analysis evaluation

producer’s design principles

consumer’s design principles

59 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

60 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

61 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

62 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

63 of 119

Weight for analyst i’s allocation to “data matching”

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

64 of 119

Weights across all

principles sum to 1

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

65 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

66 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

67 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

68 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

69 of 119

the mean

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

70 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

71 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

72 of 119

Mean weight for principle k

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

73 of 119

Relative to a baseline principle b

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

74 of 119

Some field-specific mean

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

75 of 119

analyst-specific deviation from the field-specific mean

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

76 of 119

vector of analysis specific resources (budget, time, etc)

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

77 of 119

vector of coefficients that indicate how each resource is related to the up- or down-weighting of principle k

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

78 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

79 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

80 of 119

the impact of the analyst consumer analytic negotiation

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

81 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

82 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

83 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

84 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

85 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

86 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

87 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

88 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

89 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

90 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

91 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

92 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

93 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

94 of 119

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

95 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

96 of 119

Strong pairwise alignment

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

97 of 119

Weak pairwise alignment

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

98 of 119

Potential pairwise alignment

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

99 of 119

Potential pairwise alignment

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

100 of 119

Potential pairwise alignment

Dropped the individual difference from field-specific mean

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

101 of 119

Potential pairwise alignment

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

102 of 119

Case Study

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

103 of 119

Improving potential pairwise alignment

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

104 of 119

Improving potential pairwise alignment

choose an analyst from your field

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

105 of 119

Improving potential pairwise alignment

agree on resources dedicated to the analysis

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

106 of 119

Improving potential pairwise alignment

have a discussion between analyst and consumer

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

107 of 119

Next steps

  • Using this framework to build generative agents leveraging large language models
  • Running large-scale simulations to observe analytic negotiations between analyst agents and client agents
  • Quantify the variation and identify strategies for successful negotiations

108 of 119

Proof of concept

  • Used GPT-4 to build two agents, Analyst and Client
  • Gave an initial prompt indicating the design principle allocations for each agent
  • Allowed the agents to interact without human input
  • Client agent determined when the analysis was “complete”
  • Asked the agents to report how their design principles changed after the interaction

109 of 119

Proof of concept

  • The analysis task was to estimate the number of statisticians in Forsyth County, North Carolina.
  • During the analysis and negotiation, the Analyst outlined a methodology considering the population’s educational attainment and applied national workforce statistics to estimate statisticians in the county.
  • After some discussion, the Analyst provided an estimate of approximately 48 statisticians.

110 of 119

Client prompt

Analyst prompt

Pretend you are a client in need of a data analysis to be completed. You would like a data analyst to complete an analysis to approximate the total number of statisticians living in Forsyth County, North Carolina. You have a finite set of resources, of which you can dedicate to the following 6 design principles of data analysis, as defined in the attached paper:

* data matching: 20%

* exhaustive: 5%

* skepticism: 5%

* second order: 5%

* clarity: 45%

* reproducibility: 20%

Do not write out your design principle weights, just keep them in mind when responding. The analyst may have their own set of principle weights in mind -- you may negotiate with the analyst to help them complete the analysis appropriately. When they are finished, state: "Thank you, you have completed the task". Do not tell them they are finished until (1) the analysis meets your expectations in terms of design principles and (2) they give you a concrete number that answers your question. You are the client, your lines will begin with Client:. Ready the attached paper and then complete the following to tell the analyst that you would like them to approximate the total number of statisticians living in Forsyth County, North Carolina:

Client:

Pretend you are a data analyst. You are tasked with a specific data analysis that the client will give you in the following prompt. You have a finite set of resources, of which you can dedicate to the following 6 design principles of data analysis, as defined in the attached paper:

* data matching: 20%

* exhaustive: 45%

* skepticism: 5%

* second order: 5%

* clarity: 5%

* reproducibility: 20%

You are going to complete the analysis for the client keeping this allocation in mind. You may negotiate with the client to take their needs into account as well. Do not write out your principle weights, just keep these in mind when answering the future questions. Read the attached paper. You are the analyst, your lines will begin with Analyst:.

Client: I would like you to conduct a data analysis to approximate the total number of statisticians living in Forsyth County, North Carolina. Please ensure the analysis adheres to certain design principles, keeping in mind the importance of clarity and reproducibility, among others. Let me know your approach and the results once you have completed the task.

Analyst:

111 of 119

112 of 119

113 of 119

Client prioritized a less exhaustive analysis at baseline

Analyst prioritized a more exhaustive analysis

114 of 119

Post negotiation, they both moved closer together

115 of 119

Analyst prioritized a less clear analysis at baseline

Client prioritized a more clear analysis

116 of 119

Post negotiation, they both moved closer together

117 of 119

design principles

for data analysis

data matching

How well does the available data match the data needed to investigate a question?

exhaustive

Are specific questions addressed using multiple, complementary methods, tooling or workflows?

skeptical

Are multiple, related explanations considered using the same data?

second-order

Does the analysis include anything that does not directly address the primary question, but gives important context to the analysis?

clarity

Does the analysis summarize data in a way that is influential in explaining how the underlying data connects to the conclusions?

reproducible

Could someone who is not the original producer take the published code and data and compute the same results?

D’Agostino McGowan, Peng & Hicks (2022)

DOI: 10.1080/10618600.2022.2104290

118 of 119

Analyst

Consumer

Analytic Negotiation

Analyst

Consumer

Baseline

Negotiation

Resolution

D'Agostino McGowan, Peng & Hicks (2023).

arXiv e-prints, arXiv-2312.

119 of 119

Design Thinking: Empirical Evidence for Six Principles of Data Analysis

Lucy D’Agostino McGowan

Wake Forest University

@LucyStats

lucymcgowan.com/talk

analyticdesigntheory.org

Presentation template by Slidesgo

Icons by Flaticon

Infographics by Freepik

Images created by Freepik