Analyzing data saves lives, II

Transcription

1 Analyzing data saves lives, II Data Science 101 Stanford University, Department of Statistics

2 link to source A more contemporary example 1997 study on the use of cell phones and car collisions, one of the first of its kind

3 Background Motor vehicle collisions are a leading cause of death in North America. Popularity of cellular phones has increased dramaticallynumber of subscribers exceeded the birth rate in the U.S in In 1997, cell phones were not widely used and many were fixed (in car) phones. Phone calls are a distraction: is there a link with collisions? Cell phones increase driver s state-of-mind No epidemic of collisions. There are many other distractions when driving

4 Some previous work Case-control study of drivers in 1979: cell phone owners vs non-owners Marginally lower rate of collisions in owners. But there are confounding factors: people who buy a cell phone may differ from others in important ways. Randomized trial using a driving simulator (1995): Worse reaction time using cell phone (2.2ms vs 1.6ms). But may not represent real driving experience.

5 1997 Study design 5890 drivers approached in collision reporting centers in Toronto, who were involved in collisions causing significant property damage but no personal injury between July 1, 1994, and August 31, had cell phones, and 742 consented to be in our study. 699 provided accurate telephone numbers that could be linked to detailed billing records. Some basic information (age, sex... ) was obtained by questionnaire Using questionnaire, police reports, and 911 calls, we carefully deduced the actual time of collision and their phone usage in the previous week.

6 1997 Study: case-crossover design Self matching case-control design: each person acts as their own control. Contrast a time period on the day of the collision with a comparable period on a day preceding the collision Case crossover analysis would identify an increase in risk if there were more telephone calls immediately before the collision than would be expected solely as a result of chance

7 Hazard and Control Windows

8 Main findings On: On the phone Results Day before/on Day before/not on Day of collision: On Day of collision: Not on Is this evidence that using cell phones when driving is dangerous?

9 Hypothetical findings Results Day before/on Day before/not on Day of collision: On Day of collision: Not on What if the findings where this?

10 Hypothetical findings Results Day before/on Day before/not on Day of collision: On Day of collision: Not on Another alternative set of findings... How can we compare these scenarios?

11 Relative risk Ratio of the probability of having a collision when using a cellular telephone at any time during a 10-minute interval the probability of having a collision when not using a cellular telephone at any time during a 10-minute interval. To estimate the Relative risk Number of concurrent cases that were not exposed the day before RR = Number of non concurrent cases that were exposed the day before

12 Looking at the data celldata <- read.csv("./data/celldatanew.csv") attach(celldata) table(dayof.on,daybef.on) daybef.on dayof.on rrisk<-sum(dayof.on*daybef.noton)/ (sum(dayof.noton*daybef.on)) rrisk [1]

13 Let s look a bit more at the data table(dayof.on,daybef.on,gender),, gender = f daybef.on dayof.on ,, gender = m daybef.on dayof.on

14 Let s look a bit more at the data table(dayof.on,daybef.on,type.of.cell),, type.of.cell = handheld daybef.on dayof.on ,, type.of.cell = handsfree daybef.on dayof.on

15 Graphical displays Cellphone usage No Yes Yes Comparison day No Collision day

16 A problem in the design You cannot be involved in a car collision if you are not driving. The authors estimated that only 65% of the subjects were driving at the same time the day before the collision. They modified the relative risk estimate, by dividing the denominator of the ratio by 0.65 rrisk<-sum(dayof.on*daybef.noton)/ (sum(dayof.noton*daybef.on)/.65) rrisk [1]

17 Better control windows What if the day before the accident was a Sunday and driving behavior is different on week-ends?

18 Results The New England Journal of Medi Relative Risk of a Collision Day before Workday Weekday Comparison Day Figure 1. Relative Risk of a Collision for Different Control Periods. Maximaluse day Matching day TABLE 2. RELAT IN 10-MINU CHARACTERISTIC All subjects Age (yr) Sex Male Female High-school graduati Yes No Type of job Professional Other Driving experience (y

19 Confidence of the estimates The values we are reporting are estimates: just one guess of the true relative risk based on the data we observed. Even if there is really a difference in your risk of having a car collision if you are on the phone or not, the variability of the sample observations might give us different estimates. We will learn how we can not only summarize the information in the data with regard to the dimension that interests us, but also how we can quantify the strenght of that information.

20 Some additional comparisons ASSOCIATION BETWEEN CELLULAR-TELEPHONE CALLS AND MOTOR VEHICLE COLLISIONS Relative Risk of a Collision Relative Risk of a Collision Time before Collision (min) Figure 2. Time of Cellular-Telephone Call in Relation to the Relative Risk of a Collision. Each minute before the collision was assessed as an independent hazard interval, with these intervals grouped in five-minute periods. Cellular-telephone activity for each hazard interval was evaluated in relation to the same period on the day before the collision. Relative risks greater than 1 indicate an association between telephone use and collisions. The vertical lines 0.1 Morning Afternoon Evening Other Time of Day Monday Tuesday Wednesday Thursday Friday Weekend Day of Week Figure 3. Consistency of Relative Risks Obtained from Different Collision Times. The graph shows estimates of relative risk for collisions at different times of the day and of the week. Morning was defined as 8 a.m. to 11:59 a.m., afternoon as noon to 3:59 p.m., evening

21 The aftermath Lots of TV, radio, newspaper press. Difficult to convey ideas of probability and risk to press and general public. Eg. relative risk of 4 is about equal to that for driving at the legal limit of alcohol. Authors quoted as saying Using cell phones in the car is as bad as driving drunk. But the exposure shorter and hence cumulative risk less. There are now many laws limiting the use of cell phones while driving.

22 Main recommendations Avoid unnecessary calls. Keep conversations brief. Suspend dialogue during hazards. In 1997: Our study may serve as a warning: if cell phones become standard equipment in cars, collision rates may increase But smart phones and texting were coming!