Commuting and health in Cambridge

,

I’m taking part in a study on commuting and health in Cambridge run by the MRC Epidemiology Unit at the Institute of Metabolic Science at Addenbrooke’s Hospital.

My participation in the study involves wearing an ActiGraph GT3X activity monitor for a week to see how much exercise I get. (In passing, $335 seems like a lot of money for a device which basically consists of a three-axis solid state accelerometer and 4 MB of memory on a 3 cm PCB. I guess that’s low production volumes for you, and maybe the rigours of medical certification: a Wii Remote is a much more sophisticated piece of kit, and that retails for $40.)

It will be interesting to see (when or if the researchers eventually publish) how they’ll deal with two obvious sources of bias. First, their sample is (at least partly) self-selected (they recruit participants who sign up at their website; I don’t know what other avenues of recruitment they have), and these participants will differ from the general population in lots of ways, not all of which will be easy to control for. It seems quite likely that people who volunteer for a study on activity will be more active than the population in general. Second, the problem of reactivity: people change their behaviour when they are being observed. It seems certain that someone whose activity is being monitored will undertake more activity than they normally would. For example, if I weren’t taking part in the study I might have taken more notice of the weather forecast and not done any cycling this afternoon. However, this second source of bias applies to all participants in this study (and indeed in any study that monitors activity), so it probably doesn’t invalidate comparisons; it’s just the absolute numbers that are affected.

Azara Blog points out a third source of bias:

What is the point of this survey? The obvious conclusion will be that people who cycle or the half dozen people who walk to work are healthier than those who drive or take a bus. What a surprise. The eventual report will probably miss the fact that there is a difference between correlation and causation. So for example, for people who work in Cambridge, richer people generally live closer to their workplace and are more likely to cycle, and of course richer people are generally healthier.

The particular causative relation alleged here (richer people choose to live closer to their workplace) may or may not be true, but it illustrates a potential source of bias: commuting distance is not an independent variable because peoples choices about where to live and work may be correlated with their choice of commuting mode.

However, the blogger goes on to suggest that the researchers may be motivated to ignore this source of bias in order to reach politically correct conclusions:

But the way the conclusion of the study will be pushed is that if only the peasants could be forced to cycle 10 miles, or walk 5 miles across a muddy field, to get to work, then the world would be a better place. It is unfortunate that at a time when the UK research councils are going to be forced to tighten their belts, that money is diverted from real research to this kind of pointless academic middle class exercise.


David Ogilvie
MRC Epidemiology Unit
Institute of Metabolic Science
Box 285
Addenbrooke’s Hospital
Cambridge
CB2 0QQ

2009-07-24

Dear Dr Ogilvie,

Research study on commuting and health in Cambridge

I was recruited to take part in this study, and I’ve spent the past week wearing the activity monitor you sent me.

I’m looking forward to seeing the results of the study when they are published. You say in the information sheet that you “will send [me] a summary of the overall results.” I’d be grateful if you could send me a copy of the complete paper; or, if you publish in an open access journal, a link to the published paper.

I will be particularly interested to see how you will handle three statistical problems:

  1. Selection bias. Your sample of participants is (at least partly) self-selected: you recruit participants who sign up at your website, and these participants will differ from the general population. In particular, the fact that they learned about the study in the first place suggests they have a particular interest in commuting and health, which may correlate with the variables you are investigating.

    It also seems quite possible that people who volunteer to have their activity monitored will be more active than the population in general. Maybe you have additional sources of participants (or reliable other studies) that will allow you to control for this bias.

  2. The observer effect (reactivity): people change their behaviour when they are being observed. It’s possible that someone whose activity is being monitored will undertake more activity than they normally would. (Indeed, I had this effect in mind when I signed up for the study: I thought it might motivate me to do a bit more exercise than I normally do.) But maybe this doesn’t affect comparisons between subgroups within the study?

  3. Determining the direction of causation (if any). Among the variables in your study are commmuting distance, commuting mode, and health. But it’s not clear which of these are dependent on which others: it’s easy to hypothesize causal connections in multiple directions between these variables.

    For example, mode may be caused by distance (walking is only suitable for short distances), but distance may be caused by mode (if you prefer a particular commuting mode, you may be able to choose the locations of work and home accordingly). Health may be caused by mode (because walking and cycling improve your fitness), but mode may be caused by health (if your health is poor, you may not be able to walk or cycle far enough or fast enough to commute). The dominant direction of causation may differ between subgroups in your study.

    There may also be hidden variables that correlate with the study variables. For example, in a commentary on your study, the “Azara Blog” hypothesized that “richer people generally live closer to their workplace and are more likely to cycle, and of course richer people are generally healthier.” If the hypothesis is true (I have no idea if it is), this would be problematic for your study because your survey doesn’t capture wealth or income. (Maybe postcode would be a good enough proxy for these variables?)

Yours sincerely,

Gareth Rees


9 September 2009

Dear Gareth

Research study on commuting and health in Cambridge

Many thanks for your letter and for being willing to take part in our study. We will be publishing the results in open-access journals where possible, and we intend to summarise the emerging findings in a series of newsletters over the next few years. We can certainly provide copies of full papers if you are interested.

I agree that people who volunteer to take part in research may systematically differ from those who do not and that measurement itself may influence behaviour. However, as you say, such potential biases are unlikely to differ systematically between subgroups. This study is more concerned with making comparisons within the study population than, for example, estimating the population prevalence of cycling (for which other data sources are available) and we can examine the influence of measurement on behaviour in this and other related studies. You also suggest that wealthier people may be more likely to live close to their workplace, to cycle to work, and to be healthier in general. I agree that this is a plausible hypothesis, and we intend to explore and take account of it in this study. People are often reluctant to answer survey questions on personal or household income, but as you say, other markers of socioeconomic circumstances are available.

The need to understand the direction of any putative causal relationships is the main reason for adopting a longitudinal design for this study. Most studies in this field rely on cross-sectional datasets from which, as you point out, it is difficult to infer evidence of causality. We will have the opportunity to examine changes over a two year period, during which some participants will move house, change job, decide to take up cycling, or fine themselves presented with new transport options. We will therefore be able to examine the temporal relationships between changes in circumstanced and changes in behaviour, and between changes in behaviour and changes in certain measures of health. Unlike the author of the Azara blog to which you refer, we have not decided what the results and conclusions will be in advance of having collected and analysed the data.

I hope you have enjoyed taking part in the study. I would be happy to discuss any of these issues further if you are interested.

Yours sincerely

Dr David Ogilvie


2010-02-27

Dear Dr Ogilvie,

Research study on commuting and health in Cambridge

I took part in this study in July of last year, and we exchanged letters about possible confounding factors in the study.

I’m writing again to alert you to a possible problem with the data in the study. I’ve just received a volunteer feedback letter which includes a graph showing the “total number of minutes of physical activity of at least moderate intensity (for example, brisk walking) recorded each day.”

I’m going to pick on just one day from the study, 2009-07-19 (there are similar problem on the other days, but of smaller magnitude). According to your graph, I did about 35 minutes of moderate intensity on that day. But according to my own diary, I cycled 100 miles that day. So in fact I did at least 8 hours (480 minutes) of moderate intensity exercise on that day.

What’s the explanation? Is it the case that the ActiGraph activity monitor simply can’t identify time spent cycling? I was instructed to attach the monitor to my waist: in cycling one generally tries to keep the upper body as still as possible, except in the highest intensity sprints and climbs. So the only accelerations registered by the monitor would have been the bumps from the road surface, which may be indistinguishable from the similar bumps experienced by a passenger in a motor vehicle.

If this guess is right, and the activity monitor ignores time spent cycling, this is going to be a bit of a problem for your study. In particular, the more time someone spends cycling, the smaller the amount of exercise they will appear to get! (Because cycling displaces other forms of activity, especially walking.)

Yours sincerely,

Gareth Rees


Subject: Actigraphs and cycling
Date: Tue, 2 Mar 2010 14:00:28 -0000
From: David Ogilvie
To: Gareth Rees <gareth.rees@pobox.com>

Dear Gareth

Many thanks for your letter about the mismatch between your Actigraph record and the cycling activity recorded in your diary. You’re absolutely right that a waist-mounted accelerometer does not ascertain cycling effectively; this is a well-known limitation of the method. In the next round of data collection (later this year) we’ll be offering the opportunity to wear more sophisticated devices (combined heart rate and movement sensors and GPS receivers) to address this problem. I hope you’ll be able to take part in that.

Best wishes
David