Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
propensity score (PS), to compare high versus low dosing of IVCS describing a bias in which the indication to allocate subjects to
in pediatric ASC. The PS method is a powerful statistical tool either groups (ie, treating with high- or low-dose steroids) is related
to control for confounding-by-indication bias in retrospective to the outcome in ways other than the comparison of interest. In our
longitudinal cohorts, but requires large datasets. We therefore case, physicians may prefer higher steroids dose in the most severe
established the largest cohort to date of 283 children with ASC end of the spectrum, those who previously failed standard dosing, or
treated with IVCS. those with prolonged oral steroid course before admission. These
factors, among others, may affect outcome of treatment, indepen-
METHODS dently of the prescribed dose. This bias may be addressed by the PS
This study used existing datasets of the prospective Outcome matching that allows controlling for many background covariates
of Steroid therapy in Colitis Individuals (OSCI) study (n ¼ 128 from simultaneously by matching on a single scalar variable. The PS in
5 centers in North America) and the retrospective OSCI study our case is the probability that a child receives at admission either
(n ¼ 99 from Toronto), both of which reported the 1-year outcome high- or low-IVCS dose, given the observed covariates (17–28). By
of children admitted for ASC (2,12). In addition, 54 children were matching pairs of children with similar probabilities of being treated
added for the present study by chart review (from Jerusalem and with high or low IVCS dose, and assuming no hidden bias, we
Liverpool) using the same design and dataset structure as the OSCI created a quasi-randomized experiment.
studies. The Jerusalem cohort was added as a convenience sample, Another method for controlling for the confounding-by-
and Liverpool was chosen because of the local general, but not indication bias is the ‘‘doubly robust’’ weighted regression
universal, tendency of prescribing higher IVCS doses. (29,30). We used this model to validate our findings using the
Data acquisition was based on the same detailed protocol in entire cohort in regression models weighted by the inverse variance
all cohorts. This includes similar eligibility criteria, case report of the PS and adjusted for the unbalanced baseline variables: age,
forms, timing of visits, and outcome. Explicit demographic, prior use of thiopurines, days of bloody stools before admission,
clinical, and laboratory data were recorded at admission, at percent of weight loss during the month before admission, PUCAI,
3 and 5 days thereafter, at introduction of second line therapy, at albumin, C-reactive protein (CRP), platelets, and ESR. In addition,
discharge and 1 year after admission. Disease activity was measured we forced into the models 3 clinically important variables: number
using both subjective physician global assessment (PGA) using of steroid courses in the year before admission, weight, and PGA
100 mm visual analogue scale, and the Pediatric UC Activity Index upon admission.
(PUCAI), a valid 6-item clinical measure of UC activity. The The explicit stages of matching and analysis with PS is
responsiveness of the PUCAI to rapid change is high, allowing described elsewhere (19). Briefly, a logistic regression model
capture of change in disease activity day by day (2,12–14). was constructed with high or low-dose IVCS as the dependent
Children were managed according to the discretion of the physician, variable (repeated using the 3 cutoff values defined above) and the
including corticosteroids dosing. Therefore, there was a significant following 17 pretreatment variables as the explanatory variables:
dosing variability in the cohorts, an essential asset to be used in the sex, age, disease duration, prior 5-aminosalycilic acid treatment,
PS matching to allow for dose-effect assessment. days of oral steroids before admission, days of bloody stools before
Inclusion criteria were children with confirmed UC by admission, number of steroid courses in the year before admission,
accepted criteria (15) at the age of 2 to 18 years, admitted for number of admissions in the year before the index admission,
IVCS treatment for ASC. In order to avoid repeated measures of the disease activity at baseline (PUCAI score and PGA), temperature
same patients, only the first admission of each child was included. at baseline, vomiting at baseline, and laboratory values at baseline
Children with infectious gastroenteritis or other positive bacterial including CRP, ESR, albumin, platelets, and hemoglobin. Next,
cultures were excluded. Hydrocortisone and other types of corti- according to the beta estimates of each variable, a PS was calculated
costeroids were standardized as methylprednisolone-equivalent for each child according to his/her individual parameters. This
using established conversion factors (16). model is based only on baseline parameters and is simply a means
The cohort was divided 3 times into ‘‘high’’ and ‘‘low’’ dose by which to assign a score for estimating the probability that a
groups, based on 3 different cutoffs, justified from different prac- subject receives either dose within our specific cohort. Overfitting is
tices (all of the doses are calculated based on the children weight on thus of less concern because the model is not intended to predict
admission): outcomes nor to be applied on unseen data.
1 1
Each child from the high-dose subgroup was individually
1. 1 mg kg day methylprednisolone to 40 mg day1, matched to a low-dose child according to the nearest value of the
referred herein as the ‘‘low cutoff’’; logit of the PS, in a blinded fashion to all outcome variables (19),
1 1 1
2. 1.25 mg kg day methylprednisolone to 50 mg day , within a caliper size of a quarter of a standard deviation, according
referred herein as the ‘‘standard cutoff’’; to Rosenbaum and Rubin (31). The goodness of the PS matching
1 1
3. 2 mg kg day methylprednisolone to 80 mg day1, was evaluated by the degree that these and other baseline covariates
referred herein as the ‘‘high cutoff’’. were balanced between the groups.
In order to test the hypothesis that children treated with oral
The primary outcome was percentage of children failing prednisone before the admission may require higher steroid doses
IVCS, based on the need for salvage therapy by hospital discharge because of downregulation of receptors, we performed a sensitivity
(colectomy or second-line medical therapy with infliximab or analysis in which only children that were treated with prednisone
calcineurin inhibitors). Secondary outcomes included percentage upon admission were included in the PS-matched analysis.
of children who had at most mild disease at day 5 (defined by Data are presented as means standard deviation or medians
PUCAI < 35 points), length of admission, and the need for salvage (interquartile range (IQR)), as appropriate. Paired data of matched
therapy during the subsequent year. children were compared using McNemar test (categorical) and
paired Student t test (continuous) and the signed-rank test as
Statistical Analysis appropriate for the distribution normality. Unpaired data were
compared using x2 test, student t test, or Wilcoxon rank sum test,
A simple comparison of the steroid dosing is not possible as appropriate. Imputation was performed using the hot deck
because of confounding-by-indication bias. The latter is a term method for parameters with <10% missing data occurring at
www.jpgn.org 59
random (32). No outcome measures were missing. All of the dL (IQR 0.12–1.62) vs 0.82 (0.24–2.42); P ¼ 0.62), ESR (32 mm/
comparisons were made using 2-sided significance levels of hour (12–45) vs 31(15–50); P ¼ 0.92), hemoglobin (10.25 g/dL
P < 0.05. Statistical analyses were performed using SPSS version (8.5–11.3) vs 10.1 (8.7–11.4); P ¼ 0.42), albumin (31 g/L (28–36)
22.0 (IBM SPSS Statistics, Armonk, NY). The study was approved vs 31(28–35); P ¼ 0.58), and platelets (450 103/mL (383–576) vs
by the institutional review board of each participating center. 449 (393–575); P ¼ 0.57, respectively.
60 www.jpgn.org
www.jpgn.org
TABLE 1. Baseline characteristics of the included cohort
Matched <1.25 mg kg1 day1 Matched 1.25 mg kg1 day1 Matched <2 mg kg1 day1 Matched 2 mg kg1 day1
Entire cohort and <50 mg day1 or 50 mg day1 and <80 mg day1 or 80 mg day1
(n ¼ 283) (n ¼ 109) (n ¼ 109) P (n ¼ 43) (n ¼ 43) P
Disease duration, mo 2 (0–14) 2.1 (0–14) 1 (0–12) 0.94 0.1 (0–3.5) 0.7 (0–10.5) 0.07
Steroid courses in 0 (0–1) 0 (0–1) 0 (0–1) 0.85 0 (0–1) 0 (0–1) 0.59
previous year
Aminosalicylates use 76 (27%) 29 (27%) 27 (25%) 0.87 3 (7%) 9 (21%) 0.11
at admission
Prednisone before 0 (0–14) 0 (0–15) 0 (0–12.5) 0.7 0 (0–13) 0 (0–13) 0.82
admission, days
Bloody stools before 21 (10–56) 21 (9–39) 21 (13–56) 0.11 20 (10–56) 30 (15–75) 0.15
admission, days
Height, cm 148 25 147 23 148 27 0.77 146 28 144 26 0.87
Weight, kg 42.8 17.5 40.6 14.5 43.0 18.7 0.28 43.2 22.1 38.9 17.8 0.33
Weight loss in the 5.9 (1.9–8.6) 6.2 (3.0–8.8) 6.4 (3.1–9.1) 0.68 6.4 (0.9–8.3) 7.5 (3.4–9.7) 0.9
month before
admission, %
Disease activity— 74 12 74 13 74 13 0.8 72 16 74 11 0.34
PGA (0–100 mm)
PUCAI 69 13 70 11 70.4 14 0.98 65 13 66 16 0.66
Moderate ( 60) 73 (26%) 22 (20%) 27 (25%) 16 (37%) 14 (32%)
Severe (65) 210 (74%) 87 (80%) 82 (75%) 27 (63%) 29 (67%)
Temperature, oC 37.3 0.8 37.3 0.6 37.3 0.9 0.84 37.3 1.2 37.2 0.8 0.47
Albumin, g/L 32.9 6.5 32.8 5.8 33.1 6.7 0.72 32.0 6.8 33.0 7.5 0.51
C-reactive protein, 1.6 (0.9–2.5) 1.25 (0.6–2.5) 1.9 (1.0–2.5) 0.28 1.75 (1–3.35) 1.28 (0.8–2.5) 0.2
mg/dL
ESR, mm/h 40 (26–56) 40 (28–60) 41 (24–59) 0.73 41 (30–56) 45 (21–58) 0.67
Hemoglobin, g/dL 10.3 2.2 10.2 2.1 10.3 2.3 0.69 10.0 2.2 10.2 2.3 0.71
61
Choshen et al JPGN Volume 63, Number 1, July 2016
35 50
30
40
25
30
N
20
N
15 20
10
10
5
0 0
0 0.5 1 1.5 2 >2.5 0 10 20 30 40 50 60 70 80 90 >100
Methylprednisone (mg/kg) Methylprednisone (mg/day)
FIGURE 1. Steroid dosing of 283 children admitted for acute colitis. A, IVCS dose for children weighing 40 kg or less (18 children were treated with
doses >2.5 mg kg1 day1 [2.7–31 mg kg1 day1]). B, IVCS dose for children >40 kg (14 with doses 100 mg day1 [100–1000
mg day1]). IVCS ¼ intravenous corticosteroids.
demonstrated the need for salvage therapy by 1 year of 11% In the PS-matched sensitivity analysis, performed on the
(reported in a seperate manuscript in this issue). children that were treated with prednisone upon admission, no
significant difference was found in the outcomes of the high- versus
Entire Cohort low-dose subgroups in all of the 3 aforementioned cutoffs (data not
presented), supporting the notion that high dose is not required also
We used the ‘‘doubly robust’’ PS-weighting technique on the in those who failed oral steroids.
entire cohort, using the standard cutoff, the unbalanced baseline
variables, and the 3 clinically important variables. The only sig-
nificantly different outcome was the length of admission with a DISCUSSION
clinically irrelevant effect size (median of 9 days (IQR 5–16) in the We assembled the largest cohort to date of children with ASC
low-dose group vs 10 days (6–21) in the high-dose group, to compare the effectiveness of different IVCS doses using a robust
P ¼ 0.005). The need for salvage therapy during admission statistical method across several outcomes and subgroups. We were
(OR ¼ 1.0 (95% CI ¼ 0.6–1.5); P ¼ 0.92), rates of PUCAI < 35 successful in balancing 25 possible confounders, approximating
points at day 5 (OR ¼ 1.2 (0.8–1.8); P ¼ 0.42), and the need for random allocation. For most outcomes and analyses, we could not
salvage therapy during the year after admission (OR ¼ 1.3 (0.9– find superiority of the high doses, including the need for salvage
2.0); P ¼ 0.18) were similar. therapy during admission, mean admission duration, and the need
80 80
Day 5 PUCAI
60 60
40 40
20 20
0 0
62 www.jpgn.org
15
meta-regression of cohort studies that encompassed 1991 patients
did not demonstrate a dose-colectomy association or advantage of
doses of methylprednisolone-equivalent >60 mg day1 (34). A
prospective trial assessed 128 children (included also in this report)
10
and found that the severity of disease on admission predicted IVCS
responsiveness rather than the administered dose (2,12). Glucocor-
ticoid bioassay did not predict response to corticosteroids in
0 pediatric ASC (35). A randomized controlled trial of 66 adults
admitted for ASC showed equivalency between continuous and
bolus IVCS dosing (36). A double-blind randomized trial compar-
0 ing outcomes of 35 adult patients treated with either adrenocorti-
0.2 0.4 0.6 0.8 1 cotropic hormone or hydrocortisone failed to identify significant
Propensity score differences in response (37). Finally, an outpatient randomized trial
FIGURE 3. Distribution of the PSs of the low- and high-dose groups in compared 3 doses of oral prednisolone and found that 40 and
the standard cutoff (ie, 1.25 mg kg day1 to 50 mg day1), show- 60 mg day1 were as effective, and superior to 20 mg day1 (38).
ing the matching zone. Children with similar PS were treated with The anecdotal evidence for the effectiveness of very high-
different doses and were therefore good candidates for matching. dose pulse steroids of up to 1000 mg day1 is limited to case
reports and is inconsistent (9–11) (39). A prospective study of 20
for salvage therapy during the year following admission. This was patients treated with pulse steroids showed remission rates of only
evident for matched children treated with doses lower or higher than 60%, similar to the rates found in local historical controls (10). A
1 mg kg1 day1 and 1.25 mg kg1 day1. Similarly, in a retrospective Japanese report found that pulse dosing in 20 children
controlled analysis of the entire cohort, high-dose steroids were achieved remission faster than in 17 children treated with standard
not associated with any of the 5 outcomes explored. dosing; however, treatment was successful in all 37 children and
Although some outcomes were better in the higher doses none required salvage therapy (9–11). A national survey from
when considering a cutoff of 2 mg kg1 day1 (or 80 mg day1), Japan showed 55% short-term remission among 21 children treated
it is likely this was influenced by further differences in local with pulse steroid therapy; the comparative results of the other 41
practice protocols in one center. For instance, the more frequent children treated with the traditional dosing are not reported (11).
60
P = 0.68 Matched-lower than
60
1.25 mg/kg/day
50 mg/day (n = 109)
50 Matched-higher than
P = 0.22 1.25 mg/kg/day 50
50 mg/day (n = 109)
40
40
Number of days
P = 0.88
%
30 30
20 20
P = 0.39
10 10
0
0
Salvage therapy Moderate-severe Salvage therapy on the Admission days
during admission disease at day 5 year following admission
FIGURE 4. Outcomes of the low- and high-dose matched groups according to the standard cutoff (1.25 mg kg day1 to 50 mg day1
methylprednisolone).
www.jpgn.org 63
Taken together, the present anecdotal evidence does not support 16. Mager DE, Lin SX, Blum RA, et al. Dose equivalency evaluation of
superiority of pulse steroids over the known
70% response to major corticosteroids: pharmacokinetics and cell trafficking and cortisol
standard dosing [11]. dynamics. J Clin Pharmacol 2003;43:1216–27.
The nonsignificant differences in the outcomes of predni- 17. Austin PC. A critical appraisal of propensity-score matching in
sone-treated children that received high- or low-dose IVCS may the medical literature between 1996 and 2003. Stat Med 2008;27:
indicate that also these children do not require higher IVCS doses. 2037–49.
18. Brookhart MA, Schneeweiss S, Rothman KJ, et al. Variable selection for
Data for this large cohort were collected in a variety of propensity score models. Am J Epidemiol 2006;163:1149–56.
hospitals in North America, Europe, and Israel, including a wide 19. D’Agostino RB Jr. Propensity score methods for bias reduction in the
spectrum of patients and health systems. The variety is an essential comparison of a treatment to a non-randomized control group. Stat Med
asset when using PS. This is, however, still a post hoc analysis 1998;17:2265–81.
study, and we cannot exclude the possibility that unknown con- 20. Fitzmaurice G. Confounding: propensity score adjustment. Nutrition
founding variables were missed. Given the paucity of children 2006;22:1214–6.
treated with doses <1 mg kg1 day1, we could not subdivide 21. Hullsiek KH, Louis TA. Propensity score modeling strategies for the
this group. Common wisdom, however, supports a minimal dose of causal analysis of observational data. Biostatistics 2002;3:179–93.
1 mg kg1 day1 up to 40 mg daily. Finally, we did not collect 22. Leon AC, Hedeker D. Quantile stratification based on a misspecified
propensity score in longitudinal treatment effectiveness analyses of
data on adverse events as dictated by the original OSCI datasets. ordinal doses. Comput Stat Data Anal 2007;51:6114–22.
Nonetheless, higher steroid doses are associated with more adverse 23. Oakes JM, Church TR. Invited commentary: advancing propensity score
events (40), although these depend on the duration of treatment, and methods in epidemiology. Am J Epidemiol 2007;165:1119–21.
are mostly reversible. Despite the inherent limitations, this study 24. Senn S, Graf E, Caputo A. Stratification for the propensity score
provides for the first time a comprehensive approach on a large compared with linear regression techniques to assess the effect of
cohort of children to explore the dose-effect relation of IVCS in treatment or exposure. Stat Med 2007;26:5529–44.
pediatric ASC. The consistent results across several analyses sup- 25. Stukel TA, Fisher ES, Wennberg DE, et al. Analysis of observational
port the use of lower IVCS doses, as dictated by the present studies in the presence of treatment selection bias: effects of invasive
cardiac management on AMI survival using propensity score and
guidelines (1–1.5 mg kg1 day1 up to 40–60 mg day1)
instrumental variable methods. JAMA 2007;297:278–85.
(33). Additional studies are, however, needed to determine the best 26. Sturmer T, Joshi M, Glynn RJ, et al. A review of the application of
dose with more certainty. propensity score methods yielded increasing use, advantages in specific
settings, but not substantially different estimates compared with con-
REFERENCES ventional multivariable methods. J Clin Epidemiol 2006;59:437–47.
1. Griffiths AM. Specificities of inflammatory bowel disease in childhood. 27. Sturmer T, Schneeweiss S, Rothman KJ, et al. Performance of propen-
Best Pract Res Clin Gastroenterol 2004;18:509–23. sity score calibration—a simulation study. Am J Epidemiol 2007;165:
2. Turner D, Walsh CM, Benchimol EI, et al. Severe paediatric ulcerative 1110–8.
colitis: incidence, outcomes and optimal timing for second-line therapy. 28. Yue LQ. Statistical and regulatory issues with the application of
Gut 2008;57:331–8. propensity score analysis to nonrandomized medical device clinical
3. Chakravarty BJ. Predictors and the rate of medical treatment failure in studies. J Biopharm Stat 2007;17:1–13.
ulcerative colitis. Am J Gastroenterol 1993;88:852–5. 29. Bang H, Robins JM. Doubly robust estimation in missing data and
4. Truelove SC, Jewell DP. Intensive intravenous regimen for severe causal inference models. Biometrics 2005;61:962–73.
attacks of ulcerative colitis. Lancet 1974;1:1067–70. 30. Kang JD, Schafer JL. Demystifying double robustness: a comparison of
5. Turner D, Griffiths AM. Acute severe ulcerative colitis in children: a alternative strategies for estimating a population mean from incomplete
systematic review. Inflamm Bowel Dis 2011;17:440–9. data. Stat Sci 2007:523–39.
6. Travis SPL, Stange EF, Lémann M, et al. European evidence-based 31. Rosenbaum PR, Rubin DB. Constructing a control group using multi-
consensus on the management of ulcerative colitis: current manage- variate matched sampling methods that incorporate the propensity
ment. J Crohn Colitis 2008;2:24–62. score. Am Stat 1985;39:33–8.
7. Turner D, Travis SP, Griffiths AM, et al. Consensus for managing acute 32. Streiner DL. The case of the missing data: methods of dealing with
severe ulcerative colitis in children: a systematic review and joint dropouts and other research vagaries. Can J Psychiatry 2002;47:68–
statement from ECCO, ESPGHAN, and the Porto IBD Working Group 75.
of ESPGHAN. Am J Gastroenterol 2011;106:574–88. 33. Turner D, Travis SP, Griffiths AM, et al. Consensus for managing acute
8. Turner D, Levine A, Escher JC, et al. Management of pediatric severe ulcerative colitis in children: a systematic review and joint
ulcerative colitis: joint ECCO and ESPGHAN evidence-based consen- statement from ECCO, ESPGHAN, and the Porto IBD Working Group
sus guidelines. J Pediatr Gastroenterol Nutr 2012;55:340–61. of ESPGHAN. Am J Gastroenterol 2011;106:574–88.
9. Kudo T, Nagata S, Ohtani K, et al. Pulse steroids as induction therapy for 34. Turner D, Walsh CM, Steinhart AH, et al. Response to corticosteroids
children with ulcerative colitis. Pediatr Int 2011;53:974–9. in severe ulcerative colitis: a systematic review of the literature and a
10. Rosenberg W, Ireland A, Jewell DP. High-dose methylprednisolone in meta-regression. Clin Gastroenterol Hepatol 2007;5:103–10.
the treatment of active ulcerative colitis. J Clin Gastroenterol 1990; 35. Turner D, Kolho KL, Mack DR, et al. Glucocorticoid bioactivity does
12:40–1. not predict response to steroid therapy in severe pediatric ulcerative
11. Nagata S, Shimizu T, Kudo T, et al. Efficacy and safety of pulse steroid colitis. Inflamm Bowel Dis 2010;16:469–73.
therapy in Japanese pediatric patients with ulcerative colitis: a survey of 36. Bossa F, Fiorella S, Caruso N, et al. Continuous infusion versus bolus
the Japanese Society for Pediatric Inflammatory Bowel Disease. Diges- administration of steroids in severe attacks of ulcerative colitis: a
tion 2010;81:188–92. randomized, double-blind trial. Am J Gastroenterol 2007;102:601–8.
12. Turner D, Mack D, Leleiko N, et al. Severe pediatric ulcerative colitis: a 37. Meyers S, Lerer PK, Feuer EJ, et al. Predicting the outcome of corticoid
prospective multicenter study of outcomes and predictors of response. therapy for acute ulcerative colitis: results of a prospective, randomized,
Gastroenterology 2010;138:2282–91. double-blind clinical trial. J Clin Gastroenterol 1987;9:50–4.
13. Turner D, Otley AR, Mack D, et al. Development, validation, and 38. Baron J, Connell A, Kanaghinis T, et al. Out-patient treatment of
evaluation of a pediatric ulcerative colitis activity index: a prospective ulcerative colitis. Br Med J 1962;2:441.
multicenter study. Gastroenterology 2007;133:423–32. 39. Sachar DB. Pulse steroids for ulcerative colitis: good news, bad news,
14. Turner D, Hyams J, Markowitz J, et al. Appraisal of the pediatric ulcera- and no news. J Clin Gastroenterol 2002;35:291–2.
tive colitis activity index (PUCAI). Inflamm Bowel Dis 2009;15:1218–23. 40. Lichtenstein GR, Abreu MT, Cohen R, et al. American Gastroenterolo-
15. Levine A, Koletzko S, Turner D, et al. ESPGHAN revised porto criteria gical Association Institute technical review on corticosteroids, immuno-
for the diagnosis of inflammatory bowel disease in children and modulators, and infliximab in inflammatory bowel disease. Gastro-
adolescents. J Pediatr Gastroenterol Nutr 2014;58:795–806. enterology 2006;130:940–87.
64 www.jpgn.org