Mostly what we’re left with is a convenience sample of some kind, usually determined by introductions from the survey workers themselves. It is absolutely the worst way to run a survey and the data is usually crap, but, worse yet, unverifiable crap.
Ideally, in a household level survey, we’d run in establish target areas for sampling, do a complete census on target areas and then perhaps take a random sample within those areas. At the minimum this would be a relatively decent approach.
Unfortunately, I often encounter one of two situations. The first is the convenience sample I mentioned above, which is inherently biased toward the social connections and thus the demographic of the survey workers themselves. If you want to do a sample of someone’s friends and family, this might be a good start, otherwise its completely awful.
The second is the “school based survey,” a design I think I hate more than all others. This travesty of sample design depends on the good graces of families which send their children to school, being lucky that the kids you are interested in show up to school the day of the survey and reasonable connections with school administrations. Worse yet, if you’re doing a survey on health, the chances that you’ll the kids you’re really interested in at school is really low. People love this awful design because it’s convenient, cheap, can be done in a short time and has the added benefit of providing one with warm feelings.
I’ve resolved myself to do neither of these again. As the manager of a Health and Demographic Surveillance System based in Kenya which monitors more than 100,000 people in two regions of Kenya, I decided I have a unique opportunity to do something a little more interesting.
In gearing up for a pilot survey to improve measurement of socio-economic status in developing country contexts, I realized that I had an incredible set of resources at my disposal. I have a full sampling frame on two sets of 50,000 people in two areas of Kenya, basic demographic information and a competent staff with sufficient time to do a project which otherwise would interfere with their regular duties.
With some help from a friend (well, much more than a friend), I maneuvered the basic of complex survey design and came up with something that might work relatively well for my purposes.
The DSS of the area of Kwale, Kenya I’m working in is divided into nine areas, each delegated to a single field interviewer who visits each of the households three times a year. Each field interviewer area is then divided into a number of subgrids, the number of which arbitrarily follows the population surveyed and the logistics of the survey rounds. Some areas are easier to survey than others. Each grid then has a number of households within them, the number of which varies depending on population density.
I want to target three areas, each of which ostensibly will represent different levels of economic development, but in reality represent different types of economic activities and lifestyles. One is relatively urbanized, another is purely agricultural and the third is occupied by agro-pastoralists who keep larger herds of large animals.
I then decided to choose 20 grids in each area at random, and then want to select up to 10 households from each selected grid again at random. The reason for choosing this strategy was purely a logistic one. Survey workers can do about 10 households in a day and I’ve given them a month (20 working days) to do it before they have to start on their next round of regular duties. Normally, I’d like to do something fancier, but without any previous data on the variables I’m interested in, it just wasn’t possible.
I have discovered that this design is called a stratified two stage cluster design which makes it all sound fancier that I really believe it to be. The advantage to using this design is that I’m able to control for the selection probabilities, which can bias the results when doing statistical tests. I have no doubt that the piss poor strategies I’ve used in the past and the dreaded “school based survey” I mentioned above are horribly biased and don’t really tell us a whole lot about whatever it is we’re trying to find out.
I used the survey package in R to determine the selection probabilities and, as I suspected, found that the probability of selection is not uniform across the sampling frame. Some households are more likely to be included in the survey, biasing the data in favor of, for example, people in more densely populated areas.
Alright, enough for now….
I was just screwing around with some data we collected a bit ago. In a nutshell, I’m working to try to improve the way we measure household wealth in developing countries. For the past 15 years, researchers have relied on a composite index based up easily observable household assets (a la Filmer/Pritchett, 2001).
Enumerators enter a households and quickly note the type of house construction, toilet facilities and the presence of things like radios, TVs, cars, bicycles, etc. Principal Components Analysis (PCA) is then used to create a single continuous measure of household wealth, which is then often broken into quantiles to somewhat appeal to our sense of class and privilege or lack thereof.
It’s a quick and dirty measure that’s almost universally used in large surveys in developing countries. It is the standard for quantifying wealth for Measure DHS, a USAID funded group which does large surveys in developing countries everywhere.
First, I take major issue with the use of PCA to create the composite. PCA assumes that inputs are continuous and normally distributed but the elements of the asset index are often dichotomous (yes/no) or categorical. Further, PCA is extremely sensitive to variations in the level of normality of the elements used, so that results will vary wildly depending on whether you induce normality in your variables or not.
It’s silly to use PCA on this kind of data, but people do it anyway and feel good about it. I’m sure that some of the reason for this is the inclusion of PCA in SPSS (why would anyone ever use SPSS (or PASW or whatever it is now)? a question for another day…)
So… we collected some data. I created a 220 question survey which asked questions typical of the DHS surveys, in addition to non-sensitive questions on household expenditures, income sources, non-observable assets like land and access to banking services and financial activity.
The DHS focuses exclusive on material assets mainly out of convenience, but also of the assumption that assets held today represent purchases in the past, which can act as pretty rough indicators of household income. So I started there and collected what they collect in addition to all my other stuff.
This time, however, I abandoned PCA and opted for Multiple Correspondence Analysis, a technique similar to PCA but intended for categorical data. The end result is similar. You get a set of weights for each item, which (in this case) are then tallied up to create a single continuous measure of wealth (or something like it) for each household in the data set.
Like PCA, the results are somewhat weak. The method only captured about 12% of the variation in the data set, which sort of begs the question as to what is happening with the other 88%. However, we got a cool graph which you can see up on the left. If you look closely, you can see that the variables used tend to follow an intuitive gradient of wealth, running from people who don’t have anything at all and shit in the shrubs to people who have cars and flush toilets.
We surveyed three areas, representing differing levels of development. Looking at how wealth varies by area. we can see that there is one very poor area, which very little variation to the others which have somewhat more spread, and a mean level of wealth that is considerably higher. All of this agreed with intuition.
“Area A” is known to be very rural, isolated and quite poor. Areas B and C are somewhat better though they are somewhat different contextually.
My biggest question, though, was whether a purely asset based index can truly represent a household’s financial status. I wondered if whether large expenditures on things like school fees and health care might actually depress the amount of money available to buy material items.
Thus, we also collected data on common expenditures such schools fees and health care, but also on weekly purchases of cell phone airtime. Interetinngly we found that over all the two were positively correlated with one another, suggesting that higher expenses do not depress the ability for households to make purchases, but found that this relationship does not hold among very poor households.
There is nothing to suggest that high expenses are having a negative effect on material assets among extremely poor households located in Area A at all. It might be the case that there is no relationship at all. This could indicate something else. Though overall there might not be a depressive effect of health care and school costs on material purchases, they might be preventing households from improving their situation. It might only be after a certain point that the two diverge from one another and households are then able to handle paying for both effectively.
Also of interest were the similar patterns found in the three areas.
In my seminal paper, “Distance to health services influences insecticide-treated net possession and use among six to 59 month-old children in Malawi,” I indicated that Euclidean (straight line) measures of distance were just as good as more complicated, network based measures.
I didn’t include the graph showing how correlated the two were, but I wish I had and I can’t find it here my computer.
Every time I’ve done presentations of research of the association of distances to various things and health outcomes, someone inevitably asks why I didn’t use a more complex measure of actual travel paths. The idea is that no one walks in a straight line anywhere, but rather follows a road network, or even utilizes a number of transportation options which might be lost in a simple measure.
I always respond that a straight line distance is as good as any other when investigating relationships on a coarse scale. Inevitably, audiences are never convinced.
A new paper came out today, “Methods to measure potential spatial access to delivery care in low- and middle-income countries: a case study in rural Ghana” which compared the Euclidean measure with a number of more complex measurements.
The conclusion confirmed what I already knew, that the Euclidean measure is just as good in most cases, and the pain and cost of producing sexy and complicated ways of calculating distance just isn’t worth it.
It’s a pretty decent paper, but I wish they had put some graphs in to illustrate their points. It would be good to see exactly where the measures disagree.
Access to skilled attendance at childbirth is crucial to reduce maternal and newborn mortality. Several different measures of geographic access are used concurrently in public health research, with the assumption that sophisticated methods are generally better. Most of the evidence for this assumption comes from methodological comparisons in high-income countries. We compare different measures of travel impedance in a case study in Ghana’s Brong Ahafo region to determine if straight-line distance can be an adequate proxy for access to delivery care in certain low- and middle-income country (LMIC) settings.
We created a geospatial database, mapping population location in both compounds and village centroids, service locations for all health facilities offering delivery care, land-cover and a detailed road network. Six different measures were used to calculate travel impedance to health facilities (straight-line distance, network distance, network travel time and raster travel time, the latter two both mechanized and non-mechanized). The measures were compared using Spearman rank correlation coefficients, absolute differences, and the percentage of the same facilities identified as closest. We used logistic regression with robust standard errors to model the association of the different measures with health facility use for delivery in 9,306 births.
Non-mechanized measures were highly correlated with each other, and identified the same facilities as closest for approximately 80% of villages. Measures calculated from compounds identified the same closest facility as measures from village centroids for over 85% of births. For 90% of births, the aggregation error from using village centroids instead of compound locations was less than 35 minutes and less than 1.12 km. All non-mechanized measures showed an inverse association with facility use of similar magnitude, an approximately 67% reduction in odds of facility delivery per standard deviation increase in each measure (OR = 0.33).
Different data models and population locations produced comparable results in our case study, thus demonstrating that straight-line distance can be reasonably used as a proxy for potential spatial access in certain LMIC settings. The cost of obtaining individually geocoded population location and sophisticated measures of travel impedance should be weighed against the gain in accuracy.
The UN keeps data on migrations patterns around the world, tracking origin and destination countries and number of migrants (Trends in International Migrant Stock: Migrants by Destination and Origin). I took some time out and created this network visualization of origin and destination countries from 2010. Other years were available, but this is all I had time for.
The size of each node represents the number of countries from which migrants arrive. By far, the most connected country is the United States, accepting more people from more countries than any other place on the planet. Most areas of the network represent geographic regions. Note that Africa is clustered at the top, and pacific island countries are clustered at the bottom.
An interesting result is that countries tend to send migrants to other countries which are only slightly better off than they are. For example, Malawi sends most of its migrants to Zambia and Mozambique, and Zambians go to South Africa, whereas those countries do not reciprocate to countries poorer than them. Wealthy countries tend to be more cosmopolitan in their acceptance of migrants.
Click on the picture to explore a larger version of the graphic.
Policy makers in the US and Europe seized on the paper as proof that cutting stimulus and social programs was a good idea, and proceeded to do so with abandon. Of course, right wingers wanted to cut money to social programs anyway, and would have done so regardless, but the paper was held out as scientific proof that it was a solid plan of action.
I won’t comment on how strange it was that Republicans were interested in science at all, given recent efforts to politicize the NSF and micromanage the grant decision process.
The trouble was that the results presented in RR were shown to be based on the selective use of data. Thomas Herndon, a 28-year-old graduate student, obtained the dataset from RR themselves and couldn’t reproduce the results.
In fact, he found that the only way to accurately reproduce the results in RR’s paper that showed that high debt restrained economic growth was to exclude important cases. When including the missing data, high debt was associated with consistently positive growth, though modestly slowed.
Originally, I took the view that this was a case of sloppy science. RR had a dataset, got some results which fit the narrative they were pushing and didn’t pursue the matter any further. Reading Herndon’s paper, however, I changed my mind.Herdon took the data and did what any analyst would do when starting exploratory analysis, he plotted it (see figure on the right). Debt to GDP ratios and growth are both continuous measures. We can do a simple scatterplot and see if there’s any evidence that would suggest that the two things are related.
To me, this is a pretty fuzzy result. Though the loess curve (an interpolation method to illustrate trend) suggest that there is *some* decline in growth overall, I’d still ding any intro stats student for trying to suggest that there’s any relationship at all. There is no way that RR, both trained PhD’s and likely having the help of a paid research assistant, didn’t produce such a plot.
Noting that the loess curve drops past approximately 120%, I calculated the median growth for each country represented. Only 7 countries have had debt to GDP ratios greater than 120% in the past 60+ years: Australia, Belgium, Canada, Japan, New Zealand, the UK and the United States. Out of these only two had (median) negative growth: Belgium (-.69%, effectively zero) and the United States (-10.94%), which has only had a debt to GDP greater than 120% one time. All other countries has positive growth under high debt, even beleaguered Japan. New Zealand can even claim a strong 9.8% growth under high debt. The US, then, is a major outlier, possibly bringing the entire curve down.
As this doesn’t fit their story, RR’s solution was to categorize debt to GDP ratios into five rough classifications, and calculate the mean growth within each group. This is a common trick to extract results from bad data. It’s highly tempting for researchers (and epidemiologists do it far too often), but a bad idea to present it without all the caveats and warnings that should go with it.
I’m not surprised that ideologues such as RR would be so keen to produce the result they did. After all, they published the popular economics work “This Time Is Different: Eight Centuries of Financial Folly” where they try to suggest that budget policy of the US in 2013 should somehow be informed by the economy of 14th century Spain.
I am, however, surprised that reviewers let this pass. If I would have been a reviewer, I would have:
1) pointed out the problems of categorization, where data doesn’t require it
2) noted that categorizing the data (or even plotting it) tears out temporal correlation. For example, one data point from 2008 (stimulus) may be put in the high debt category, but another from 2007 (crash) in the low debt category. While budgets of one year may have little to do with the budget of another, the economy of one year is likely related to the economy of the previous year.
3) questioned the causal mechanisms behind debt and growth. This is obviously a deep question for economists (and not epidemiologists), but of particular import. When does the economy start to react to debt? I’m pretty sure that there is a lag effect as spending bills tend to space disbursements over the course of the fiscal year.
The RR debacle should be a lesson, not only to economists, but to all scientists. While we may always be under pressure to produce results and hope that those results fit and support whatever position we take, shoddy methods don’t get us off the hook. In RR’s case, I would call this fabrication. A good many studies are merely guilty of wishful thinking, but the chance always exists that someone will come out of the woodwork and expose our flaws. After all, that’s what science is all about.
A couple of weeks ago, I attended a lecture on network analysis where the investigators analyzed popular political books on Amazon.com.
Amazon lists not only information on the book but also the titles, in order of purchasing frequency, of other books that customers may have purchased. The researchers here were able to identify left leaning and right leaning books by examining the purchasing habits of Amazon customers.
Decibel “is America’s only monthly extreme music magazine” and has been in publication since 2004. Every year, they publish the titles of the 40 best metal records of the year, according to their review staff.
Here is 2012’s list:
40 Gojira – L’Enfant Sauvage
39 Meshuggah – Koloss
38 Agalloch – Faustian Echoes EP
37 The Shrine – Primitive Blast
36 Incantation – Vanquish In Vengeance
35 Samothrace – Reverence To Stone
34 Devin Townsend Project – Epicloud
33 Panopticon – Kentucky
32 Saint Vitus – LILLIE: F-65
31 Mutilation Rites – Empyrean
30 Author & Punisher – Urus Americanus
29 A Life Once Lost – Ecstatic Trance
28 Asphyx – Deathhammer
27 Farsot – Insects
26 Gaza – No Absolute For Human Suffering
25 Inverloch – Dark/Subside
24 Swans – The Seer
23 Horrendous – The Chills
22 Killing Joke – MMXII
21 Early Graves – Red Horse
20 Liberteer – Better To Die On Your Feet Than Live On Your Knees
19 High On Fire – De Vermis Mysteriis
18 Napalm Death – Utiltarian
17 Torche – Harmonicraft
16 Grave – Endless Procession Of Souls
15 Satan’s Wrath – Galloping Blasphemy
14 Testament – Dark Roots Of Earth
13 Cattle Decapitation – Monolith Of Inhumanity
12 Blut Aus Nord – 777: Cosmosophy
11 Municipal Waste – The Fatal Feast
10 Pig Destroyer – Book Burner
09 Paradise Lost – Tragic Idol
08 Royal Thunder – CVI
07 Enslaved – Riitiir
06 Neurosis – Honor Found In Decay
05 Pallbearer – Sorrow and Extinction
04 Witchcraft – Legend
03 Evoken – Atra Mors
02 Baroness – Yellow & Green
01 Converge – All We Love We Leave Behind
I looked all of these records on Amazon. For each of them, I noted which of the others were in the first 12 titles that were purchased with it, creating a 40 by 40 adjacency matrix where rows (i) and columns (j) represented records. For each entry, a zero was noted where the customer which purchased the i-th record did not purchase the j-th record, and a one where they did.
I found that many of the records on the list were purchased with one another. The most common record purchased in combination with another on the list was Neurosis‘ “Honor Found in Decay.” Fifteen of the other records on this Top 40 were purchased with “Honor Found in Decay.”
In network terms, the Degree of this record would be 15. Pallbearer’s “Sorrow and Extinction” had a degree of 11, Royal Thunder’s “CVI” and Blut Aus Nord’s “777: Cosmosophy” both had a degree of 9.
The network of Decibel’s Top 40 looks like this:
You can see that some records get purchased with other records more than others. The size of the dots represent the degree of the record.
Now, I did some cluster analysis on the data, looking for related groups of records within the network. Using R, I produced the following dendrogram:
There are two major clusters, each with its own subcluster (dendrograms are hierarchical). One includes Converge, Neurosis, Pallbearer Royal Thunder, Evoken and Inverloch with a subcluster including only the first four. These are all bands that might be expected to be purchased with one another. The other big one includes all the rest. Main clusters are designated by color.
I found one containing the three entries for Baroness, Municipal Waste and Napalm Death, very different bands. I’m truly not sure why those three would be in a cluster together (is the cluster is based on lonliness in the network?).
Anyway, I’m done, but glad I got any results at all. I’ll let readers (especially metal fans!) interpret the results.