Causal Inference: predic1on, explana1on, and interven1on

Size: px

Start display at page:

Download "Causal Inference: predic1on, explana1on, and interven1on"

Monica Nichols
5 years ago
Views:

1 Causal Inference: predic1on, explana1on, and interven1on Lecture : Mechanisms, Interven1ons, and Randomized Controlled Trials Samantha Kleinberg samantha.kleinberg@stevens.edu

2 Blame Who s at fault for a car accident? Why was Alice late to a mee1ng?

3 The vice- president of a company went to the chairman of the board and said: We are thinking of star1ng a new program. It will help us increase profits, but it will also harm the environment. The chairman of the board answered: I don t care at all about harming the environment. I just want to make as much profit as I can. Let s start the new program. They started the new program. Sure enough, the environment was harmed.

4 The vice- president of a company went to the chairman of the board and said: We are thinking of star1ng a new program. It will help us increase profits, and it will also help the environment. The chairman of the board answered: I don t care at all about helping the environment. I just want to make as much profit as I can. Let s start the new program. They started the new program. Sure enough, the environment was helped.

5 Knobe effect/side effect effect Study: 78 people in NYC park, each one randomly assigned to one story Harm: chairman blamed for harming environment (82%), ac1ons deemed inten1onal Help: chairman given no credit for helping environment (77%), ac1ons deemed uninten1onal difference p<.001 Knobe, J. (2003). Inten1onal Ac1on and Side Effects in Ordinary Language. Analysis, 63,

6 Inten1onality Goals vs side effects Intent vs actual outcome

8 Norm viola1on No credit for doing what one should Not lifering Not chea1ng on exams Blame if viola1ng behavioral standard Jaywalking

9 Students in a dorm get a copy of their biology final. Most students cheat. John Granger did not. The exam is graded such that only the top 20 students get an A. Granger was 20 th. A pre- med student was 21 st and missed the GPA cutoff or med school by.07 points! Did Granger cause her to fail to meet the cutoff? Is he to blame? Were his ac1ons good/bad? Alicke, M. D., Rose, D., & Bloom, D. (2011). Causa1on, norm viola1on, and culpable control. The Journal of Philosophy, 108(12),

11 Who are the subjects in these studies?

13 Morris & Peng Fundamental afribu1on error Ul1mate afribu1on error

14 Visual percep1on

15 Newspaper reports What s key advantage of this vs asking for explana1ons?

16 Evalua1ng explana1ons & counterfactuals

explain why it was done Miller JG (1984) Culture and the Development of

19 Afribu1on differences/age Par1cipants: age 8, 11, 15 and adult from India and U.S. Subjects were asked to give example of a good behavior and bad one, and explain why it was done Miller JG (1984) Culture and the Development of Everyday Social Explana1on. Journal of Personality and Social Psychology 46(5):

20 Do views change over 1me? Or does understanding of expecta1ons change?

22 Revisi1ng Morris & Peng Subjects: ethnic Chinese in Hong Kong, ethnic whites in U.S. Replicated Morris & Peng fish result But perceived consensus rather than actual beliefs mediated effect Zou X, Tam K- P, Morris MW, Lee S- L, Lau IY- M, Chiu C- Y (2009) Culture as common sense: perceived consensus versus personal beliefs as mechanisms of cultural influence. Journal of Personality and Social Psychology 97(4):

23 Biases Sampling bias Confirma1on bias Seeking posi1ve evidence Discoun1ng nega1ve evidence

24 Next week: journal club Goal: evaluate and discuss work other than your own that may be outside your field, learn about new work For all papers: read in depth, and prepare to discuss For your assigned paper: Read the ar1cle (+ other references if needed for context) Prepare to give a brief summary Prepare ques1ons and lead discussion

25 Format Summary of paper What s the main argument/hypothesis? How is this supported? Overview experiments, figures Discussion Bring comments, cri1cism, ques1ons Timing min/paper

26 Papers Diabetes/sugar (Isocaloric fructose restric9on and metabolic improvement in children with obesity and metabolic syndrome) Poverty (Poverty Impedes Cogni9ve Func9on) Downsize (Invi9ng Consumers To Downsize Fast- Food Por9ons Significantly Reduces Calorie Consump9on)

27 On to new stuff..

28 Coffee Study of 402,260 AARP members Men who had 2-3 cups/day 10% less likely to die than non- coffee drinkers, Women 13% Women drinking 4-5 cups 16% less likely to die of stroke, accidents, diabetes, among others

30 Mechanisms Causal processes, causal interac1ons, and causal laws provide the mechanisms by which the world works; to understand why certain things happen, we need to see how they are produced by these mechanisms Salmon, W. C. (1984). Scien9fic explana9on and the causal structure of the world. Princeton University Press, Princeton, NJ.

31 Example Smoking causes yellowed fingers The tar in cigarefe smoke stains fingers

32 Regulari1es/Probabilis1c causality People usually develop a fever aver contrac1ng the flu Mechanisms An infec1on sends signals to the brain, which regulates body temperature, which in turn raises this in response to the infec1on.

33 Mechanisms and interven1on Smoking à LC - If we don t know how smoking causes LC, can only intervene to reduce smoking rates - If we know how it causes LC, can intervene in later parts of the process

34 What is a mechanism? Like causality, no agreement/single defini1on General idea A system of interac1ng parts that regularly produce some change

35 Defini1ons of mechanisms Glennan: A mechanism underlying a behavior is a complex system which produces that behavior by the interac1on of a number of parts according to direct causal laws. [Glennan, 52] Machamer, Darden, Craver (MDC): Mechanisms are en11es and ac1vi1es organized such that they are produc1ve of regular changes from start or set- up to finish or termina1on condi1ons [MDC, 3]

36 Key points Produc1on Invariance, predictability Set of interac1ons/ac1vi1es Not just a mechanism, a mechanism for a behavior

37 Granularity Carl Sagan (343): Some ancient Asian cosmological views are close to the idea of an infinite regression of causes, as exemplified in the following apocryphal story: A Western traveler encountering an Oriental philosopher asks him to describe the nature of the world: ``It is a great ball res1ng on the flat back of the world turtle. ``Ah yes, but what does the world turtle stand on? ``On the back of a larger turtle. ``Yes, but what does he stand on? ``A very percep1ve ques1on. But it's no use, mister; it's turtles all the way down.'' Sagan, C. (1979). Broca's brain: Reflec9ons on the romance of science. Ballan1ne Books

38 Hierarchy Image from Dhorspool at en.wikipedia hfp:// transla1on/

39 Mechanis1c causality Glennan: a rela1on between two events (other than fundamental physical events) is causal when and only when these events are connected in the appropriate way by a mechanism [Glennan, 56] MDC: An en1ty acts as a cause when it engages in a produc1ve ac1vity It is not the penicillin that causes the pneumonia to disappear, but what the penicillin does. [MDC, 6]

40 Probabilis1c causality Recall: Umbrella vendors and rain Pirates and global warming

41 Combining probability and No plausible way mechanisms Pirates can lower the earth s temperature Umbrella vendors can produce rain Two sources of evidence Neither necessary, but together can corroborate rela1onship Russo, F. and Williamson, J Interpre1ng Causality in the Health Sciences. Interna9onal Studies in the Philosophy of Science. 21(2):

42 Evidence for causality Neither difference making nor mechanisms are sufficient to establish causality Each provides different suppor1ng informa1on Plausible connec1on between cause and effect (the how) Cause actually has impact on effect Dis1nguish between causal rela1onship itself and evidence suppor1ng its existence

43 Example: gene1c varia1ons and vo1ng behavior Fowler and Dawes (2008): people with certain variants of MAOA are more likely to vote than those with others Charney and English (2012, commentary): Same variants also linked to IBS, schizophrenia, and other traits. Implausible that single variant explains all of these.

44 Example: mechanism w/o causality It is not known what causes most cases of au1sm in humans. But, know that problems w/signaling in brain may be a cause and have drugs for this. If drugs work in cases where cause not known, can postulate it was a signaling problem.

45 Structural equa1ons, counterfactuals Pearl/Woodward and mechanisms For interven1on [do] need to know altera1on doesn t affect mechanisms The structural eq. aren t affected by the interven1on When eq. invariant in this way, said to correspond to mechanisms

46 Further Reading Dowe, P. (2000). Physical causa9on. Cambridge University Press. Glennan, S. S. (1996). Mechanisms and the nature of causa1on. Erkenntnis (1975- ), 44(1), Machamer, P., Darden, L., & Craver, C. F. (2000). Thinking about mechanisms. Philosophy of Science, 67(1), Woodward, J. (2005). Making things happen: A theory of causal explana9on. Oxford University Press, USA.

47 We want to know whether a heartburn medica1on works, so we compare incidence of heartburn in users/non- users of the drug. However, people who take it are more likely to have heartburn.

48 Interven1ons to causes Running shoes A - > 5K race Running shoes B - > 5K race If pace different, what can we conclude? Could have improved in ability over 1me (or gofen injured in between) Weather may have changed (external effect of different 1me periods) Delayed effect of training during first 1me period

49 Ideal interven1on Change only the cause not any causes of it, or other causes of the effect and see if effect s1ll happens

50 Interven1on and RCTs Ideal: two groups that are iden1cal in every way except one assigned to possible cause. Any differences between the groups then must be a result of the cause.

51 RCT Two groups: control and interven1on Assignment to group is random Various methods for randomizing: e.g. by pa1ent, by site Only difference is the policy/treatment

52 Limits (more later) Tradeoff between control of study And generalizability of results

53 Not just for medicine! Webpage design Poli1cal campaigns Educa1on

54 Randomiza1on 1747: Lind discovers treatment for scurvy But assigned men on boat to dietary addi1ons (sea water, citrus, etc). Could have led to bias

55 Why randomize? Sever link between causes of interven1on and effects (selec1on bias) E.g. birth control pills and pregnancy Isolate cause Single difference between groups, removes confounding

56 Example Tes1ng text messages to promote physical ac1vity Randomize 13- year- olds in a school to either text messages promo1ng 30 minutes of ac1vity, or just giving weather report Problem: contamina1on between groups Solu1on: cluster randomiza1on

57 Who to randomize? Back to heartburn medica1on Should popula1on be? - Everyone - People with history of heartburn - People w/heartburn and not taking drugs that may interact w/proposed one - People in a certain age group

58 Rothwell, P. M. (2005). Trea1ng individuals 1 external validity of randomised controlled trials:to whom do the results of this trial apply?. Lancet, 365(9453), 82-9

59 Popula1on/pa1ents not truly random Hospital/clinic - who par1cipates? Clinician - who refers pa1ents? Pa1ent - who chooses to enroll?

60 Who to analyze? Not everyone who starts a study finishes.. May be removed by inves1gator May withdraw May be lost to follow- up Survival bias

61 Controlling 1950: Hill and others use RCT to study treatment for tuberculosis Compared streptomycin to standard of care (bed rest) Assigned treatment with numbered & sealed envelopes

62 Instead of does effect happen or not aver treatment, how does result compare to standard or no treatment Why is this needed?

63 Blinding Single: Pa1ent unaware of what treatment being received Double: Pa1ent + clinician unaware of what treatment being administered/received Triple: Double + person analyzing results not aware of which group is which

64 Example RCT of two treatments + placebo for mul1ple sclerosis Examina1on by blinded and unblinded neurologists Unblinded Neurologists showed one treatment beneficial at 6, 12, 24 months, p- value<0.05 Noseworthy, J. H., Ebers, G. C., Vandervoort, M. K., Farquhar, R. E., Ye1sir, E., & Roberts, R. (1994). The impact of blinding on the results of a randomized, placebo- controlled mul1ple sclerosis clinical trial. Neurology, 44(1),

65 Sidenote on placebos Treatment that is known to have no impact on the effect (e.g. sugar pills, fake surgery) Why can t we use no treatment? The act of treatment can induce effects

66 Placebo effect Dosage Injec1on vs pill Color of pills Packaging/branding Surgeries

67 Open placebo

69 Why blind? Confirma1on bias Placebo effect

70 Using RCT results Heartburn study was a success! Reduced frequency and severity Now: new 80 y.o. pa1ent on 10 meds. Should this new one be prescribed?

72 What s the cause here? Interven1on: office chairs Outcome: weight loss Failure in new popula1on. How? Weight loss in first test!

73 Unit: what s difference in outcome between treatments for an individual? Easy when homogeneous pop Sub- popula1on: group with par1cular feature unaffected by treatment Popula1on: distribu1on of outcomes, no conclusions about individual outcomes Holland PW, Rubin DB (1988) Causal inference in retrospec1ve studies. Evalua1on Review 12(3): 203.

74 Observa1onal study shows women taking HRT aver menopause have decreased risk of heart afacks (37% lower death rate, 53% lower risk of CV death) HERS trial: RCT showing no effect WHI: RCT where heart afacks increase 29% (from 30 to 37 per 10,000 person- years) Latest: HRT may be beneficial if it s started early

75 n=1 How can we figure out which of two interven1ons is best for one individual? Instead of randomizing people, randomize order of treatment

76 Considera1ons How many treatment periods? How to randomize order? Washout period?

77 RCT: advantages Experimental Remove confounding

78 What do RCTs tell us?

79 If an RCT shows a measurable effect of a variable, does that mean it s causal? If an RCT doesn t show an effect, does that mean the variable is not causal?

80 Evalua1ng an RCT

81 Who s being studied?

82 Groups balanced?

83 Who s doing evalua1on?

84 What s the control?

85 Sample size?

To argue from the fact that RCTs have certain advantages, other things being equal, to the claim that the RCT is a gold standard, is like arguing that since

86 To argue from the fact that RCTs have certain advantages, other things being equal, to the claim that the RCT is a gold standard, is like arguing that since being tall makes for a good high- jumper, it follows that a 6ʹ elderly drunkard with a spinal injury is bound to be a befer high- jumper than a 5ʹ 11 Olympic athlete.

87 Problems with RCTs Can be Infeasible Unethical External validity True randomiza1on difficult Cost (+ dura1on, sample size) Power

88 Infeasible/unethical There will never be an RCT to determine: - whether parachutes prevent death from sky diving - if smoking causes lung cancer - whether socioeconomic status causes health

89 External validity Goal of an RCT: test whether A works in popula1on P, use A in some other popula1on P A: BCP P: Women P : Men

90 Cost Sample size (each addi1onal person costs money and also increases recruitment 1me) Data length

91 What do we need to know before we can use the results of an RCT? Study is internally valid i.e. it can answer the ques1on it aims to answer Factors affec1ng external validity Characteris1cs of seƒng Selec1on of pa1ents Characteris1cs of pa1ents Follow- up Control? Blinding single, double, triple See also: Rothwell, P. M. (2006). Factors that can affect the external validity of randomised controlled trials. PLoS Clin Trials, 1(1), e9.

92 Going from here to there Works somewhere: X caused Y in P Works in general: X causes Y Will work here: X causes Y in P Cartwright, N., & Hardie, J. (2012). Evidence- Based policy: A prac9cal guide to doing it be[er. OUP US

93 Recall INUS Remember: cause is insufficient but necessary part of some unnecessary but sufficient condi1on Full set contains all factors such that effect will be produced Includes what we may think of as background condi1ons For external validity, need to know the background condi1ons If oxygen necessary for cause to be effec1ve, but oxygen is absent from the target case, then won t be effec1ve

94 Cartwright s main point Cause doesn t happen in a vacuum there are always necessary condi1ons for it to be effec1ve (at a minimum, things that prevent it from being effec1ve must be absent) To know that an interven1on will work, we need to know a) what these condi1ons are and b) that they re present

Health Informa.cs. Lecture 7. Samantha Kleinberg

Health Informa.cs. Lecture 7. Samantha Kleinberg Health Informa.cs Lecture 7 Samantha Kleinberg samantha.kleinberg@stevens.edu Prac.cali.es: preserving privacy Good prac.ces for preserving privacy Physical security Locked cabinet/office, don t print