Paperonium

← Back to the plain-language case

The logistic-regression "odds ratios", including those in the abstract, appear to be log-odds coefficients

Psychol Med 2022;52(14):3289-3296 (published online 18 March 2021) · doi:10.1017/S0033291721000088 · PMC9693673

Summary

The paper reports two multivariable logistic regressions for clinically elevated hsCRP (⩾3 mg/L) and labels every estimate an odds ratio (OR). One of them is negative ("OR −4.39, 95% CI −7.04 to −1.97"), and six more have negative lower confidence limits (four distinct intervals; the age and gender intervals appear in both models). An odds ratio cannot be negative. Read as log-odds coefficients, the printed values hang together: every printed CI agrees with its printed p-value on significance at 0.05 (the interval excludes 0 exactly when p < 0.05), and the paper's own Table 1 counts reproduce the estimates closely (unadjusted log-odds 2.39 against the printed 2.44). Read as ORs, the two main-effect intervals in model 2 contradict their p-values. On the coefficient reading, the abstract's headline odds ratios are understated about 2.7- to 4.7-fold: "OR 2.44" corresponds to an OR of about 11.5 (95% CI 3.6 to 51.4), "OR 1.52" to about 4.6 and "OR 1.09" to about 3.0. The direction and significance of each finding are unchanged, but the effect sizes as printed are on the wrong scale, and read literally as ORs, the CI given for depression (0.27 to 2.01) includes 1 and would contradict the stated significance.

What the paper reports

Methods, "Statistical analysis":

To address the second hypothesis, two models of multivariable logistic regression were used. The first assessed how much risk comorbid depression and overweight contributed to the hsCRP levels ⩾3 mg/L relative to other participant groups, with age and gender as covariates. The second assessed how much risk depression and overweight status (BMI ⩾ 25) separately contributed to hsCRP levels ⩾3 mg/L, with age and gender as covariates.

Results, "High sensitivity C-Reactive Protein", first logistic paragraph (model 1):

Overweight patients with depression were at a significantly increased risk of having hsCRP ⩾3 mg/L [OR 2.44, 95% confidence interval (CI) 1.28 to 3.94, p < 0.001], overweight controls did not have a significantly increased risk of hsCRP ⩾3 mg/L (OR 1.51, 95% CI −0.33 to 2.87, p = 0.13), normal-weight patients with depression did not have a significantly increased risk of hsCRP ⩾3 mg/L (OR 1.08, 95% CI −0.6 to 2.4, p = 0.28), while normal-weight controls were at a significantly decreased risk of hsCRP ⩾3 mg/L (OR −4.39, 95% CI −7.04 to −1.97, p < 0.001). Age and gender were not significantly associated with hsCRP ⩾3 mg/L [OR 0.01, 95% CI −0.04 to 0.06, p = 0.621, OR 0.61, 95% CI −0.19 to 1.47, p = 0.145, respectively].

Results, same section, second logistic paragraph (model 2):

Depression status was significantly associated with hsCRP ⩾3 mg/L (OR 1.09, 95% CI 0.27 to 2.01, p = 0.013) and overweight status was significantly associated with hsCRP ⩾3 mg/L (OR 1.52, 95% CI 0.7 to 2.41, p < 0.001). Age and gender were not significantly associated with hsCRP ⩾3 mg/L [OR 0.62, 95% CI −0.04 to 0.06, p = 0.647, OR 0.01, 95% CI −0.19 to 1.47, p = 0.137, respectively].

Abstract, "Results":

Multivariable logistic regression showed that comorbid depression and overweight significantly increased the risk of clinically elevated hsCRP levels ⩾3 mg/L (OR 2.44, 1.28–3.94). In a separate multivariable logistic regression model, overweight status contributed most to the risk of having hsCRP levels ⩾3 mg/L (OR 1.52, 0.7–2.41), while depression also contributed a significant risk (OR 1.09, 0.27–2).

Table 1, row "hsCRP ⩾3 mg/L (%)". The columns are overweight & depressed (N = 69), overweight control (N = 35), normal weight & depressed (N = 55), normal weight control (N = 57), statistic, p:

hsCRP ⩾3 mg/L (%) | 26 (37.7%) | 5 (14.3%) | 6 (10.9%) | 3 (5.3%) | x2 = 25.96 | <0.001

The conclusions that rest on these models. Abstract, "Conclusion":

Comorbid depression and overweight status are associated with increased hsCRP, and the coexistence of these conditions amplified the risk of clinically elevated hsCRP levels. Overweight status contributed most to the risk of clinically elevated hsCRP levels, but depression also contributed to a significant risk.

Discussion, first paragraph:

Of clinical relevance, overweight patients with depression had the highest risk for hsCRP levels ⩾3 mg/L, indicating clinically elevated peripheral inflammation.

What the paper's own numbers show

Method. In logistic regression a coefficient β is a log odds ratio, and the odds ratio is exp(β). An OR must be positive and its null is 1, while a coefficient can take any sign and its null is 0. I made three checks:

  1. Sign. I checked each printed estimate and CI limit for values ⩽ 0, which no OR can take.
  2. CI against p-value. I took each printed CI as a 95% Wald interval and back-calculated the two-sided p-value it implies under two readings: as a coefficient (null 0) and as an OR (null 1, computed on the log scale). I also checked, more simply, whether each interval excludes the null exactly when the printed p is below 0.05. The printed intervals are not symmetric about their estimates, and most look like profile-likelihood intervals (see check 3), so a Wald back-calculation from them is only approximate. The significance check does not depend on that assumption.
  3. Refit from Table 1. I refitted both models without the age and gender covariates, using only the grouped counts in Table 1 (40 cases with hsCRP ⩾3 mg/L among 216 participants). The paper's covariate-adjusted models need individual data, so the refits show the scale and approximate size of the estimates, not their exact values. I report both Wald and profile-likelihood intervals for the refits, because most printed intervals are asymmetric about their estimates in the way profile-likelihood intervals are (2.44 lies 1.16 above its lower limit of 1.28 and 1.50 below its upper limit of 3.94).

Every number below is printed by recompute.py.txt. Quoted passages keep the paper's minus signs (U+2212); elsewhere minus signs are typed as hyphens.

Quantity Location Reported Recomputed from the paper's numbers
hsCRP ⩾3 mg/L by group Table 1 26/69, 5/35, 6/55, 3/57; χ² = 25.96, p < 0.001 37.7%, 14.3%, 10.9%, 5.3%; Pearson χ² = 25.96, df 3, p = 9.7 × 10⁻⁶ (matches)
Overweight + depressed v. reference Abstract; Results, model 1 "OR 2.44 (1.28 to 3.94)", p < 0.001 As a coefficient, the CI implies p = 0.0003 (matches p < 0.001); as an OR, it implies p = 0.002. Unadjusted refit: log-odds 2.39 (profile 1.26 to 3.87; Wald 1.13 to 3.65), p = 0.0002. Crude OR = (26/43)/(3/54) = 10.9. If the printed value is a coefficient, OR = exp(2.44) = 11.5 (3.6 to 51.4)
Overweight controls v. reference Results, model 1 "OR 1.51 (−0.33 to 2.87)", p = 0.13 Negative lower limit is impossible for an OR. As a coefficient, the CI includes 0 (not significant, agreeing with p = 0.13), but a Wald back-calculation gives p = 0.064, about half the printed 0.13. Refit: log-odds 1.10 (profile -0.37 to 2.74; Wald -0.40 to 2.60), Wald p = 0.15. exp(printed) = 4.5 (0.72 to 17.6)
Normal-weight depressed v. reference Results, model 1 "OR 1.08 (−0.6 to 2.4)", p = 0.28 Negative lower limit is impossible for an OR. As a coefficient, the CI includes 0 (not significant, agreeing with p = 0.28), but a Wald back-calculation gives p = 0.16, about half the printed 0.28. Refit: log-odds 0.79 (profile -0.60 to 2.39; Wald -0.65 to 2.23), Wald p = 0.28. exp(printed) = 2.9 (0.55 to 11.0)
"Normal-weight controls" Results, model 1 "OR −4.39 (−7.04 to −1.97)", p < 0.001 Impossible as an OR. With four groups and an intercept, one group must be the reference and gets no coefficient of its own. Normal-weight controls are the reference implied by the other three contrasts. Their observed log-odds is ln(3/54) = -2.89, and the refit intercept is -2.89 (profile -4.30 to -1.89), so −4.39 is most consistent with the model intercept, which in the adjusted model refers to age 0
Depression Abstract; Results, model 2 "OR 1.09 (0.27 to 2.01)", p = 0.013 Read as an OR, the CI contains 1 and implies p = 0.87, contradicting p = 0.013. Read as a coefficient, it implies p = 0.014 (matches). Refit: log-odds 1.12 (profile 0.31 to 2.03), p = 0.010. exp(printed) = 3.0 (1.31 to 7.46)
Overweight Abstract; Results, model 2 "OR 1.52 (0.7 to 2.41)", p < 0.001 Read as an OR, the CI contains 1 and implies p = 0.18, contradicting p < 0.001. Read as a coefficient, it implies p = 0.0005 (matches). Refit: log-odds 1.46 (profile 0.68 to 2.32), p = 0.00045. exp(printed) = 4.6 (2.01 to 11.1)
Age, gender Results, model 1 "OR 0.01 (−0.04 to 0.06)", p = 0.621; "OR 0.61 (−0.19 to 1.47)", p = 0.145 Negative lower limits are impossible for ORs. As coefficients, implied p = 0.70 and 0.15. exp(printed) = 1.01 (0.96 to 1.06) per year and 1.84 (0.83 to 4.35)
Age, gender Results, model 2 "OR 0.62 (−0.04 to 0.06)", p = 0.647; "OR 0.01 (−0.19 to 1.47)", p = 0.137 The age estimate 0.62 lies outside its own CI (-0.04 to 0.06). The two point estimates appear to be transposed relative to their CIs and to model 1 (0.01 with -0.04 to 0.06; 0.61 with -0.19 to 1.47). Both CIs include 0 and both p-values are above 0.05, so significance still agrees on the coefficient reading
Size of understatement Abstract 2.44, 1.52, 1.09 exp(printed) / printed = 4.7×, 3.0×, 2.7×

In the refit of model 1, the printed intervals and the profile-likelihood intervals line up closely: 1.28 to 3.94 against 1.26 to 3.87, -0.33 to 2.87 against -0.37 to 2.74, and -0.6 to 2.4 against -0.60 to 2.39. For overweight + depressed, depression and overweight, the printed intervals also share the refit's upward skew (for example 1.16 below and 1.50 above the estimate, against 1.13 and 1.48 in the refit). That pattern is what R's confint() gives for a glm fit on the coefficient scale. The Methods state that all analyses used R 4.0.0.

The Wald back-calculation is less exact. On the coefficient reading it matches the printed p for overweight + depressed (0.0003 v. < 0.001), "normal-weight controls" (0.0007 v. < 0.001), depression (0.014 v. 0.013), overweight (0.0005 v. < 0.001) and model 1 gender (0.15 v. 0.145), and is close to it for model 1 age (0.70 v. 0.621). For overweight controls (0.064 v. 0.13) and normal-weight depressed (0.16 v. 0.28) it gives about half the printed value, a gap a reader rerunning the script will see. In these two rows the printed limits are close to the refit's limits, but the printed estimates (1.51, 1.08) are higher than the refit estimates (1.10, 0.79). The printed estimates therefore sit nearer the upper limit (1.84 below and 1.36 above; 1.68 below and 1.32 above). The back-calculation divides the estimate by an SE taken from the interval width, so a larger estimate with the same width gives a larger z and a smaller p-value. The printed p-values for these two rows are close to the unadjusted refit Wald p-values (0.15 and 0.28). Because the printed intervals look like asymmetric profile intervals, a Wald back-calculation from them is only approximate, and these two rows cannot be resolved further without the model output. The gap is in size only: on the coefficient reading, no printed interval contradicts its printed p-value on significance at 0.05.

Figure

Printed logistic-regression values, all labelled 'OR', on their own linear scale, with refits from Table 1

Each row is one group or main-effect estimate the paper prints with the label "OR"; the age and gender rows are not shown. Orange circles and lines are the printed estimate and 95% CI, on the linear scale on which they are printed. Blue squares and lines are log-odds coefficients refitted, without the age and gender covariates, from the hsCRP ⩾3 mg/L counts in Table 1, with profile-likelihood 95% intervals. The shaded region holds values below 0, which no odds ratio can take. The solid line at 0 is the null for a log-odds coefficient, and the dashed line at 1 is the null for an odds ratio. Whether each printed p-value (right) is below 0.05 matches whether its interval crosses 0, not whether it crosses 1. That is clearest for depression and overweight in model 2, whose intervals cross 1 but are reported as significant. The row labelled "normal-weight controls" has no blue refit because the reference group of a categorical predictor has no contrast of its own.

Printed values read as odds ratios v. exp(printed values), on a log odds-ratio axis

The same estimates on a log odds-ratio axis. Orange is the printed value taken literally as an OR. Lower limits below 0 cannot be placed on this axis and are drawn as arrows. Blue is exp(printed value), the odds ratio implied if the printed numbers are coefficients, labelled with its point estimate. The intercept-like "normal-weight controls" row is omitted because its printed value is negative. The age and gender rows of both models are not shown either.

Why it matters

The three figures in the abstract are the paper's quantitative answer to its second aim, "what extent overweight status and depression status contributed to clinically elevated hsCRP levels". As printed, they suggest modest associations: odds roughly 2.4, 1.5 and 1.1 times higher. If they are coefficients, the implied odds ratios are about 11.5, 4.6 and 3.0, a 2.7- to 4.7-fold difference. A reader citing these ORs, or pooling them in a meta-analysis, would carry the wrong magnitude. Read literally as ORs, the printed intervals also contradict the paper's own significance statements: the depression interval (0.27 to 2.01) and the overweight interval (0.7 to 2.41) both include 1. A reader who takes the label at face value could therefore conclude that neither main effect is significant, which is the opposite of what the abstract states.

The qualitative conclusions do not depend on the scale. Overweight-and-depressed participants have the highest odds, overweight contributes more than depression, and both are significant in model 2. Those orderings and significance calls are the same on either scale, and the Table 1 refits support them. The second-paragraph age and gender estimates are a separate, smaller transcription problem.

Explanations considered

Status

As of 2026-10-23:

Suggested correction

  1. State that the values reported in the logistic-regression paragraphs of the Results and in the abstract are log-odds coefficients (β), or replace them with odds ratios, exp(β), and their exponentiated CIs. If the printed values are the model coefficients, the abstract would read approximately:
  2. "OR 11.5, 95% CI 3.6–51.4" for comorbid depression and overweight;
  3. "OR 4.6, 2.0–11.1" for overweight;
  4. "OR 3.0, 1.3–7.5" for depression.

These values should be confirmed from the original model output. 2. In model 1, report the reference group explicitly (apparently normal-weight controls), and either drop the "−4.39" or label it as the intercept rather than as a risk estimate for normal-weight controls. 3. In the model 2 paragraph, check the age and gender estimates, which appear transposed (0.62 is printed with the interval -0.04 to 0.06). 4. Consider giving the full model output (coefficients, SEs, intercept and coding) in a table or supplement.