Paperonium

← Back to the plain-language case

Abstract attaches the most- and least-sitting case counts to the wrong exposure groups

BMJ (2011) · doi:10.1136/bmj.d3867 · PMC3266839

Summary

The abstract illustrates the primary association with raw counts: "41/104 720 v 16/14 565 cases in most inactive v least inactive". In Table 2, 41/104 720 belongs to score 2, the group that sat least and the reference group, and 16/14 565 belongs to score 6, the group that sat most. The abstract therefore attaches each count to the opposite extreme. Read as printed, the counts give a crude rate ratio of 0.36 for most versus least inactive, which suggests a protective pattern. With Table 2's labels the ratio is 2.81, which agrees with the hazard ratios of 2.73, 2.68 and 2.34 that the paper reports. The hazard ratios, the trend test and the conclusion are not affected; the error is in the labelling of one parenthetical in the abstract's primary result.

What the paper reports

Abstract, Results:

There was an association between time of sitting and risk of idiopathic pulmonary embolism (41/104 720 v 16/14 565 cases in most inactive v least inactive in combined data; P<0.001 for trend). The risk of pulmonary embolism was more than twofold in women who spent the most time sitting compared with those who spent the least time sitting (multivariable hazard ratio 2.34, 95% confidence interval 1.30 to 4.20).

Methods, Exposures:

Thus, each woman was assigned to one of five categories of time sitting, with point values ranging from 2 (low time sitting in 1988 and 1990) to 6 (high time sitting in 1988 and 1990).

Methods, Statistical analysis:

For all analyses of physical inactivity, we used the group with the lowest amount of time sitting as the reference group.

Table 2, header and first row ("Amount of time sitting* (score 2=lowest, 6=highest), combined data from 1988 and 1990"; columns 2, 3, 4, 5, 6):

No of cases/person years | 41/104 720 | 53/175 518 | 104/209 002 | 54/88 228 | 16/14 565 | —

Table 2, hazard ratio rows (score 2 is the reference, 1.00):

Age adjusted HR (95% CI) | 1.00 | 0.79 (0.52 to 1.19) | 1.31 (0.90 to 1.88) | 1.62 (1.08 to 2.44) | 2.73 (1.53 to 4.88) | 1.29 (1.15 to 1.45), <0.001

Multivariable HR (95% CI)‡ | 1.00 | 0.79 (0.53 to 1.19) | 1.29 (0.89 to 1.85) | 1.51 (1.01 to 2.28) | 2.34 (1.30 to 4.20) | 1.25 (1.12 to 1.40), <0.001

Results, second paragraph:

For women with the most physical inactivity, the risk of pulmonary embolism was more than twice that of women with the least physical inactivity (multivariable hazard ratio 2.68, 1.50 to 4.79).

What the paper's own numbers show

Method. I took the cases and person-years for each sitting score from Table 2 and the number of participants per score from Table 1. I checked that the totals match the text, used the Table 1 sitting-hours distribution to confirm which way the score runs, and computed crude incidence rates per 100 000 person-years. I then computed the crude rate ratio for most versus least sitting twice: once with Table 2's group labels and once with the labels the abstract gives. A crude rate ratio is not adjusted, so it need not equal the Cox hazard ratios, but it should point the same way and be of similar size. Every number below is printed by recompute.py.txt.

Quantity Location Reported Recomputed from the paper's own numbers
Total cases Abstract; Results; Table 2 268 41 + 53 + 104 + 54 + 16 = 268
Total participants Abstract; Table 1 69 950 12 451 + 20 718 + 24 549 + 10 403 + 1829 = 69 950
Direction of the score Methods; Table 1 (1988 sitting) 2 = low, 6 = high sitting score 2: 100% sat <10 h/week; score 6: 100% sat ≥41 h/week
Score 2 (least sitting, reference) Table 2, column 2 41/104 720 39.2 per 100 000 person-years
Score 6 (most sitting) Table 2, column 6 16/14 565 109.9 per 100 000 person-years
Group labelled "most inactive" Abstract 41/104 720 equals Table 2 score 2, the least-sitting group
Group labelled "least inactive" Abstract 16/14 565 equals Table 2 score 6, the most-sitting group
Crude ratio, most v least, abstract's labels Abstract (implied) 0.36 (39.2 ÷ 109.9)
Crude ratio, most v least, Table 2 labels Table 2 — 2.81 (109.9 ÷ 39.2; approximate 95% CI 1.57 to 5.00)
HR, score 6 v score 2 Table 2; Results; Abstract 2.73 (age adjusted); 2.68 and 2.34 (multivariable) consistent with 2.81, not with 0.36

The crude rates by score (39.2, 30.2, 49.8, 61.2, 109.9 per 100 000 person-years) follow the same pattern as the age-adjusted hazard ratios (1.00, 0.79, 1.31, 1.62, 2.73). Follow-up per woman is similar across scores (7.96 to 8.51 person-years), so person-time does not distort the comparison. The two counts in the abstract are correct numbers; only the group names attached to them are reversed.

Figure

The abstract's two counts match Table 2's extreme columns, but with the group labels reversed

Figure 1. Left panel: the parenthetical in the abstract's Results, with each count (orange, as printed) and the group name the abstract gives it. Right panel: the same two counts in Table 2, with the column they occupy, the group the paper's Methods and Table 2 define for that column, and the reported hazard ratio. The crude rates (blue) are computed from the printed counts. The connectors join identical counts: 41/104 720 is "most inactive" in the abstract but the least-sitting reference group in Table 2, and 16/14 565 is "least inactive" in the abstract but the most-sitting group in Table 2. Taken as printed, the abstract's comparison gives a crude most-versus-least ratio of 0.36; with Table 2's labels it is 2.81.

The crude ratio agrees with the reported hazard ratios only under Table 2's labelling

Figure 2. Most versus least sitting on one log axis, with a thin reference line at 1 (no difference). Orange squares with bars are the three hazard ratios (95% CI) for score 6 versus score 2 reported in Table 2, the Results and the abstract. Blue circles are crude rate ratios computed from the Table 2 counts: the open circle uses the abstract's labels (0.36), the filled circle uses Table 2's labels (2.81). Only the filled circle lies within the reported confidence intervals; the open circle lies on the other side of the null. The log scale is used because the quantities are ratios; it shows 0.36 and 2.81 as equally far from 1, which they are (one is the reciprocal of the other).

Why it matters

Explanations considered

Status

Suggested correction

In the abstract's Results, replace

(41/104 720 v 16/14 565 cases in most inactive v least inactive in combined data; P<0.001 for trend)

with

(16/14 565 v 41/104 720 cases in most inactive v least inactive in combined data; P<0.001 for trend)

so that the counts match Table 2: score 6 (most sitting) 16 cases in 14 565 person-years, about 110 per 100 000 person-years, and score 2 (least sitting) 41 cases in 104 720 person-years, about 39 per 100 000 person-years. Keeping the order of the counts and swapping the labels ("in least inactive v most inactive") would be an equivalent fix. No other part of the paper needs to change.