Please enable JavaScript.
Coggle requires JavaScript to display documents.
Good & Bad Stats - Coggle Diagram
Good & Bad Stats
-
-
Communicating Risk
Communicating Risk\n- One of biggest implications of health/clin research\n- One of biggest media misconceptions when communicating risk not adequately reporting baselines\n - Relative vs absolute risk\n- But is still a risk so should not completely dismiss due to manipulative stats
-
Relative Risk\n- Measure of risk to develop condition comparing groups\n - Group that engages w. act\n - Group w/out. act\n- Assume one group have 0 risk\n - But always underlying stat & risk for developing in both groups
-
Processed Meat & Cancer (Gallagher, 2015)\n- Consuming 50g processed meat a day increases colorectal cancer chance by 18%
Limitation\n- Represents relative risk\n - Measure of developing cancer in bacon vs non-bacon eaters\n - But rate of cancer not 0 in non-bacon eaters
-
Problematic Social Media Use & Depression (Shensa et al., )\n- Associated w. 9% increase in odds of depressive symptoms in random sample of population age 19-32\n - Measured w. brief depression scale\n- Odds = Association of outcome in compared groups\n- Odds 1.09 = Outcome occur 9% more often in A than B
Media & Risk\n- Media headlines manipulative & exaggerated\n - Report relative risk as absolute risk
-
Forensic Statistics
Victims Table\n- Typically killed more women\n- Most ages between 60-90 years\n- Most occurred in early 1990s\n - But w. gap
-
Could Shipman be Caught Sooner\n- Inquiry presented signed vs expected death certificated\n - Signed death certificates vs expected mortality rate\n- O - E w. baseline 0 = Mortality\n - Larger number = Excess mortality
Accusation\n- Type 1 error = False +\n- Type 2 error = False -\n- Could perform statistical test to compare O vs E cumulative deaths\n - Would have been significant @ p=.004 in 1979
Problem\n- If done for every person in UK would be 63,000 stat tests in 2016\n- Running many statistical tests can lead to many false +\n - Make it difficult to find genuine true result
Type 1 Error Rate\n- Multiple comparisons inflate false + rate\n - When using p-vale accept 5% are false +\n- In 2016 3,150 people would have shown as sig higher mortality rates than expected
Controlling\n- Sequential testing w. accumulated data\n - Sample size not fixed\n- Eventually p-value will be sig
Sequential Probability Ratio Test (SPRT)\n- Threshold reached in 1985\n - After 40 deaths\n- Algorithm account for Alpha & Beta level
Could Be Caught Sooner\n- Alerting system piloted for GPs in UK after Shipman inquiry\n- Identified another GP w. higher mortality rates\n- Subsequent investigation found GP worked in area w. large no. retirement homes\n - Worked hard to help residents stay out of hospital for their death
Conclusion\n- Need to explore reasons as stats not always provide them\n- Even w. sophisticated algorithm & stats to help make sense of data
Data
Data
- Any piece of info that can be used for analysis/be presented as factual
- Data & stats presented on own need interpretation
- Stats speak for themselves not true
- Human involvement needed but where humans is bias
Why Need for Data Research
- Surrounded by data
- Accessibility to science & research drastically increased in 21st
- More opportunities to see data & stats
Benefit of Stats
- Highly informative & can help decision making to better world
- Form body of evidence for change
Limitation of Stats
- Highly misleading
- Can help deceive individual consuming them
What Can Go Wrong
- Data collection
- Processing & analysis
- Presentation
Going Wrong in Data collection
- Inadequate sampling
- Select specific sample as know will support what communicate
- Not be open about sampling
- Biased measures/tools
- Write q's in specific way to elicit reponse
- Leading q's
- Design scale to exploit natural tendency to respond
Processing & analysis
- Deal w. outliers
- Cherry picking
Presentation
- Lack of transparency
- Don't tell about previous wrongdoings
- Subjective interpretation
- Not consider alt explanations
Journey Into Psych Data Analysis & Stats
- Understand, analyse & use complex data
- Retrieve and organise information from different sources
- Make critical judgements and evaluations to gain different perspectives on a question
- Be sensitive to contextual and interpersonal factors, including behaviour and social interaction
- Bigger picture & what being portrayed that need further consideration
Graph Crimes
Mismatch\n- Wrong type of presentation for data\n - Implies relationship/lack of relationship where is none\n- Wrong scale
Types of Manipulation
- Wrongly & non scientifically add to graph
- Adjust/remove baseline
- Reverse scale
- Presentation of graph to imply relationship
- Pie charts should = 100%
Adjust/Remove Baseline
- Truncate Y-axis
- Gee-Whiz graphs (Huff, 1993)
Truncate Y-Axis
- Starting axis at value other than 0
Truncation
- Sometimes needed
- If relationship masked when not truncated
- Some believe line graphs should be exempt from truncation
- Problematic as must judge where appropriate
Avoiding Issue
- Most graphs should start @ zero
- Unless 0 on scale arbitrary/useless
- Unless small but important change hidden/difficult to seen when 0 baseline used
Gee-Whiz Graph (Huff, 1993)
- Focus on section of graph rather than whole
- Cut out part of graph
- Not see full picture & rest of scale
Gee Whiz Graphs (Huff, 1993)
- Primarily issue w. bar charts but easily done w. line graphs
Why Truncation Problematic
- Quick scanning of graphs the point of having them
- To support & quick analysis
- Truncated vs non-truncated figures (Yang et al., 2021)
Truncated vs Non-Truncated (Yang et al., 2021)
- Presented figures
- Asked participants to estimate how different 2 points were\n- Truncation effect observed even w. written warning & across humanities/quantitative fields
- Truncated rated/estimated as having greater difference between data points
- Some effects reduced in subject in both conditions, more in quantitative but not significantly
Defence for Truncation
- Y-axis still plainly in sight
Reverse Scale
- Up = Decrease
- Down = Increase