Accepted Manuscript Does pregnancy alter the long-term course of multiple sclerosis? Igor Karp, MD, MPH, PhD Alexandra Manganas, BCom Marie-Pierre Sylvestre, PhD Annie Ho, Elaine Roger, Pierre Duquette, MD PII:
S1047-2797(14)00145-8
DOI:
10.1016/j.annepidem.2014.04.007
Reference:
AEP 7645
To appear in:
Annals of Epidemiology
Received Date: 21 December 2013 Revised Date:
21 March 2014
Accepted Date: 15 April 2014
Please cite this article as: Karp I, Manganas A, Sylvestre MP, Ho A, Roger E, Duquette P, Does pregnancy alter the long-term course of multiple sclerosis?, Annals of Epidemiology (2014), doi: 10.1016/j.annepidem.2014.04.007. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Title: Does pregnancy alter the long-term course of multiple sclerosis? Authors: Igor Karp, MD, MPH, PhD,1,2 Alexandra Manganas, BCom,3 Marie-Pierre Sylvestre,
RI PT
PhD,1,2 Annie Ho,3 Elaine Roger,4 Pierre Duquette, MD4,5
Affiliations: 1 Centre de recherche du Centre hospitalier de l'Université de Montréal
2
SC
(CRCHUM), Montréal, Québec, Canada
Département de médecine sociale et préventive, Université de Montréal, Montréal,
M AN U
Québec, Canada 3
Université de Montréal, Montréal, Québec, Canada
4
MS Clinic, Hôpital Notre-Dame du CHUM, Montréal, Québec, Canada
5
Faculté de Médecine (Neurology), Université de Montréal, Montréal, Québec,
TE D
Canada
Corresponding author: Igor Karp, Department of Social and Preventive Medicine, University
EP
of Montreal, 850 rue St Denis, #S02-366, Montreal, Quebec H2X 0A9. Tel: 514-890-8000, extension 15908 Fax: 514-412-7106
AC C
Email: [email protected]
Word count (main text only): 3141 Words in abstract: 225 Number of Tables: 2 Number of Supplementary Tables: 7 Number of Figures: 3
1
ACCEPTED MANUSCRIPT
Keywords: multiple sclerosis; pregnancy; prognosis; cohort; epidemiology; longitudinal studies List of abbreviations: RRMS, relapsing-remitting multiple sclerosis; SPMS, secondary progressive multiple sclerosis; EDSS, Expanded Disability Status Scale; RR, rate ratio; CI,
AC C
EP
TE D
M AN U
SC
RI PT
confidence interval;
2
ACCEPTED MANUSCRIPT
ABSTRACT Purpose: The purpose was to examine the impact of pregnancy on the rates of relapses,
RI PT
progression to irreversible disability, and transition to secondary progressive multiple sclerosis (SPMS) in patients with relapsing-remitting multiple sclerosis (RRMS).
Methods: We retrospectively followed two subcohorts of women with RRMS: pregnant
SC
(n=254) and non-pregnant (n=423). We obtained data on demographic, life-style, and clinical characteristics from patient records. Poisson and logistic regressions estimated the rate ratios
M AN U
(RR) associated with pregnancy as a function of time. Confounding was controlled by propensity-score adjustment, and post-baseline selection bias was controlled by inverse probability weighting.
Results: In the pregnant and non-pregnant subcohorts, respectively, 300 and 787 relapses, 15
TE D
and 27 transitions to SPMS, and 11 and 34 progressions to irreversible disability were documented. Adjusted RRs (95% confidence intervals) shortly after baseline were 0.67 (0.49; 0.92) for relapses, 0.16 (0.03; 0.79) for irreversible disability, and 1.25 (0.39; 3.96) for SPMS.
EP
The corresponding estimates at 5 and 10 years were, respectively, 1.04 (0.72; 1.52), 0.82 (0.36;
AC C
1.88), and 2.33 (1.03; 5.26) and 1.62 (0.84; 3.14), 4.14 (0.89; 19.22), and 4.33 (1.10; 16.99). Conclusions: Pregnancy likely ameliorates the short-term course of RRMS in terms of the rates of relapses and progression to irreversible disability. Over the long term it appears to have no material impact on these outcomes, and might in fact accelerate the rate of transition to SPMS.
3
ACCEPTED MANUSCRIPT
BACKGROUND Relapsing-remitting multiple sclerosis (RRMS) is an immune-mediated chronic inflammatory disease that is more common in women than in men. With a typical age of onset in the 3rd or 4th
RI PT
decades, the potential impact of pregnancy on RRMS course is of practical importance. While originally pregnancy was thought to worsen the course of the disease, [1] in the last several decades an alternative hypothesis was proposed, according to which pregnancy may actually
SC
have a favorable impact on RRMS [2,3].
Several epidemiological studies have attempted to address the competing hypotheses.
M AN U
However, the interpretation of, and the inference from, these studies has been difficult due to small sample sizes and various biases, of both individual studies (notably, confounding and selection bias) and their aggregate (i.e., publication bias). [4-6] Furthermore, most previous studies assessed only short-term consequences of pregnancy, usually in terms of a single
TE D
outcome. However, of even greater medical and scientific relevance would be knowledge of both the short and long-term impact of pregnancy on the triad of patient-relevant outcomes: relapse rate, transition to secondary progressive multiple sclerosis (SPMS) and progression to
EP
irreversible disability.
We set out to produce new evidence on the long-term impact of pregnancy on the course
AC C
of RRMS.
METHODS
Study population
We used data from medical records of 1,317 female RRMS patients followed at Notre-Dame Hospital Multiple Sclerosis Clinic in Montreal, Canada, from 1977 to 2010. Because at issue is a
4
ACCEPTED MANUSCRIPT
prognostic topic – namely, the prospective course of RRMS as a function of pregnancy status – data were structured as if pregnancy were a health intervention [7]. We retrospectively formed two subcohorts: the pregnant and the non-pregnant. Specifically, the non-pregnant subcohort was
RI PT
formed at the time of clinic entry (if the woman was not pregnant at that time), with this time point representing the subcohort’s baseline. The pregnant subcohort was formed among women in whom pregnancy occurred during follow-up, with the onset of pregnancy representing the
SC
pregnant subcohort’s baseline. Thus, each woman could contribute one observation to the nonpregnant subcohort and as many observations as the number of post-clinic-entry pregnancies to
M AN U
the pregnant subcohort. The inclusion criteria at baseline in both subcohorts were: (i) confirmed diagnosis of RRMS; (ii) age between 15 and 50 years, (iii) no menopause or history of hysterectomy; (iii) no prior transition to SPMS. Furthermore, in the analyses addressing the occurrence of progression to irreversible disability, the woman was not to have experienced that
TE D
outcome prior to baseline. Study outcomes
Three outcome measures were used: relapse rate (in analysis 1), transition to SMPS (in analysis
EP
2), and progression to irreversible disability (in analysis 3). The prevailing clinical practice in the Multiple Sclerosis clinic at the Hôpital Notre-Dame du CHUM, Montréal, the diagnosis of
AC C
SPMS is typically achieved when, in a patient with the Expanded Disability Status Scale (EDSS) score of 4 or higher, there is at least a 1-unit increase in that score over a period of at least 6 months without an intervening relapse. Progression to irreversible disability was operationalized as the first documented instance of the EDSS score being ≥4 [8]. Follow-up
5
ACCEPTED MANUSCRIPT
Each patient was followed from baseline until the occurrence of one of the following events, whichever came first: (i) a given study outcome (in the analyses addressing the occurrence of non-recurrent outcomes, i.e., in analyses 2 and 3); (ii) the end of study period (i.e., July 1st,
RI PT
2010); (iii) loss to follow-up, defined as no clinic visit for more than two years; (iv) transition to SPMS (in analyses 1 and 3); (v) the onset of post-baseline pregnancy; (vi) abortion. The latter five follow-up-terminating events thus represented a censoring mechanism.
SC
Statistical analysis
Data on follow-up for each woman were broken down in 3-month intervals and pooled across the
M AN U
subjects [9]. Due to the non-experimental nature of the study, two major threats to validity had to be addressed: confounding (due to potential non-comparability of the pregnant and non-pregnant women in terms of prognostic indicators) and prospective selection bias (due to potentially nonrandom censoring). Thus, before estimation of the associations between the study outcomes and
TE D
the pregnancy status, we first fitted two kinds of regression models: one, estimating the probability of membership in the pregnant subcohort as a function of its determinants (represented by the propensity score), [10] and the other, estimating the probability of censoring
EP
[11].
AC C
Estimation of propensity scores for control of confounding The dependent variable in the logistic model was a binary indicator of pregnancy subcohort membership (1 if yes, 0 if no). The potential confounders in the propensity score model were ascertained at baseline and included: time since RRMS onset, age, number of relapses in the past three months, history of prior pregnancies, EDSS score, smoking status, history of use of hormone therapy, history of use of oral contraceptives, use of interferon beta in the past three
6
ACCEPTED MANUSCRIPT
months, and use of Copaxone in the past three months. Furthermore, we added to these “maineffects” pairwise interaction terms for all the variables and quadratic terms for continuous variables. The propensity-score model was estimated according to the stepwise selection
RI PT
procedure applied to the above set of variable.
Before proceeding to the fitting of the association models of interest, we examined the distributions of the estimated propensity scores in the pregnant and non-pregnant cohorts, which
SC
revealed that they were only partially overlapping. However, inclusion of observations from nonoverlapping regions could potentially result in non-positivity bias [11]. To avoid this problem,
M AN U
we carried out “propensity score trimming” and restricted the final analytic set to observations with propensity scores between the 5th and 95th centile values [12].
Estimation of inverse-probability-of-censoring weights for control of selection bias
TE D
During follow-up, as the cohort underwent some attrition, we used the method of inverseprobability-of-censoring weighting to minimize the possibility of prospective selection bias [11]. In both subcohorts, a multiple pooled logistic regression was estimated [9,13]. The dependent
EP
variable was the event of not being censored by the end of the given 3-month interval. The independent variables included a set of variables measured at baseline (namely, time since
AC C
RRMS onset, age, number of relapses in the past three months, history of prior pregnancies, EDSS score, smoking status, history of use of hormone therapy, history of use of oral contraceptives, use of interferon beta in the past three months, and use of Copaxone in the past three months), supplemented with time-dependent variables: time since baseline, current EDSS score, number of relapses in the past three months, use of interferon beta in the past three months, and use of Copaxone in the past three months.
7
ACCEPTED MANUSCRIPT
Unstabilized weights were estimated by computing the reciprocal of the estimated probability of being uncensored in the given interval. To stabilize the weights, we estimated an analogous pooled logistic regression model with only time-invariant covariates so that stabilized
RI PT
weights were estimated by multiplying the unstabilized weights by the probability of being
Estimation of effect of pregnancy on the study outcomes
SC
uncensored, as estimated by this model [14].
Generalized linear models weighted by the inverse of the probability of being uncensored were
M AN U
estimated: pooled Poisson regression models in analysis 1 and pooled logistic regression models in analyses 2 and 3. All the models were estimated in the framework of generalized estimating equations to account for intra-individual correlation; the working correlation matrix was used. Since the probability of the non-recurrent study outcomes in a given 3-month interval is very
TE D
low, the pooled logistic regressions emulate Cox proportional-hazards regressions in estimating the RR [13].
The independent variables in all the models included an indicator of the pregnancy status,
EP
time since baseline, an interaction term between these two variables, and the propensity score. To account for potential non-linearity of the effect of time, we considered several alternative
AC C
approaches, namely: a quadratic time function as well as regression splines with one interior knot. Based on the Pan’s Information Criterion (QIC), the linear time function was retained [15]. Since interpretation of a time-function of RRs for non-recurrent outcomes can be challenging, [16] we estimated standardized cumulative incidence curves as a function of pregnancy status and time since baseline. Specifically, we estimated the predicted conditional survival probability for each 3-month interval as a function of codeterminants’ values at baseline
8
ACCEPTED MANUSCRIPT
and time since baseline for each woman under the scenario – actual or counterfactual – of membership in the pregnant subcohort. Sequential multiplication of these probabilities produced the estimates of cumulative survival for each woman up to each time point. The cumulative
RI PT
incidence estimate for each woman at each time point was obtained by deriving the complement of the corresponding cumulative survival estimate. Averaging the cumulative incidence estimates across all the observations produced the estimate of the cumulative incidence for pregnant
incidence for non-pregnant women was obtained similarly.
SC
women standardized to the whole study population. The standardized estimate of the cumulative
M AN U
Missing values of the independent variables were dealt with by the missing-value indicator method [17]. All analyses were performed using R 2.14.1 and SAS version 9.1 for Windows.
TE D
RESULTS
Of the 1,317 women in the source population, 1,026 met the non-pregnant subcohort membership criteria. During follow-up, there were documented 290 pregnancies in 183 women
EP
who met the pregnant subcohort membership criteria. After exclusion of observations where the estimated propensity score was below the 5th centile or above the 95th centile value, the number
AC C
of observations in the restricted analytic set was 254 (contributed by 165 women) in the pregnant subcohort and 423 in the non-pregnant subcohort. A total of 1,035 women-years were accrued in the pregnant subcohort and 2,655 women-years in the non-pregnant subcohort, with the median duration of follow-up being 2.9 and 6.9 years, respectively. Distributions of relevant characteristics of the subjects at baseline are summarized in Table 1. Compared with women in the non-pregnant subcohort, women in the pregnant
9
ACCEPTED MANUSCRIPT
subcohort were less likely to have been pregnant before (78% vs 64%), have a history of cigarette smoking (28% vs 22%) and hormone therapy (14% vs 11%), and to have had at least one relapse within three months before baseline(22% vs 10%), but more likely to have a history
RI PT
of use of oral contraceptives (39% vs 55%) and to have used disease-modifying drugs (DMDs) within three months before baseline(8% vs 11% for interferon beta and 3% vs 4% for Copaxone).
SC
During follow-up, in the pregnant and non-pregnant subcohorts, respectively, 300 and 787 relapses, 15 and 27 events of transition to SPMS, and 11 and 34 events of progression to
M AN U
irreversible disability were documented. The crude relapse rates, transition to SPMS, and progression to irreversible disability were, in the pregnant and non-pregnant subcohorts, 0.285 and 0.295, 0.014 and 0.010, and 0.012 and 0.016 events per person-year, respectively. Results of regression modelling for estimation of propensity scores are shown in
TE D
Supplementary Table 1. The model had a high discriminating ability and the distributions of the potential confounders between the subcohorts within categories of the propensity score quartiles were well-balanced.
EP
Results of regression modelling for estimation of censoring weights are shown in Supplementary Tables 2-7. All the models were characterized by a moderate discriminating
AC C
ability, with narrowly distributed stabilized weights, suggesting that censoring was not heavily influenced by the characteristics at issue. Results of regression models for estimation of effects of pregnancy are shown in Table 2. The estimate of the adjusted relapse RRs (95% CI) for the pregnant vs non-pregnant shortly after baseline was 0.67 (0.38, 0.65), implying that at that time, the adjusted relapse rate in the pregnant subcohort was 50% lower than in the non-pregnant. The exponentiated estimated product-term
10
ACCEPTED MANUSCRIPT
coefficient for the subcohort membership status and time since baseline was 1.09 (1.02, 1.17), implying that with each additional year of follow-up, the adjusted relapse RR increased by this factor. Consequently, the estimates of the adjusted relapse RRs at 5 and 10 years post-baseline
RI PT
were 1.04 (0.72; 1.52) and 1.62 (0.84; 3.14), respectively. Figure 1 shows that the relapse rate decreased over time in both subcohorts and that while the relapse rate in the pregnant subcohort was lower in the first several years, the difference gradually diminished to the point of
SC
convergence around 8 years after baseline.
With the transition to SPMS as the outcome, the estimated adjusted RR (95% CI) for the
M AN U
pregnant vs non-pregnant shortly after the baseline was 1.16 (0.36, 3.77) and the exponentiated estimated product-term between the pregnancy subcohort status and time since baseline was 1.14 (0.93, 1.39), implying that there was virtually no evidence that rates of transition to SPMS shortly after the baseline are different and only weak evidence that the ratio between these rates
TE D
increases over time. Accordingly, the estimated adjusted RRs (95% CIs) at 5 and 10 years postbaseline were 2.33 (1.03; 5.26) and 4.33 (1.10; 16.99), respectively. Figure 2 shows that the divergence in cumulative incidence of transition to SPMS emerged soon after the baseline and
EP
continued throughout the follow-up, with the estimated 10-year cumulative incidence rates reaching 17% and 8% in the pregnant and non-pregnant subcohorts, respectively.
AC C
With conversion to irreversible disability as the outcome, the estimate of the adjusted RR (95% CI) for the pregnant vs non-pregnant shortly was 0.16 (0.04, 0.79) and the exponentiated estimated product-term between pregnancy subcohort status and time since baseline was 1.38 (1.06, 1.80), implying that there was moderately strong evidence that the rate of conversion to irreversible disability is lower in the pregnant than in the non-pregnant shortly after the baseline and that, over time, the rate in the pregnant decreases at a slower rate than in the non-pregnant.
11
ACCEPTED MANUSCRIPT
Accordingly, the estimated adjusted RRs (95% CIs) at 5 and 10 years post-baseline were 0.82 (0.36; 1.88), and 4.14 (0.89; 19.22), respectively. Figure 3 shows that the cumulative incidence rate of this outcome in the pregnant subcohort was lower than in the non-pregnant subcohort for
RI PT
the most of the follow-up, with the difference in cumulative incidence peaking at about 5 years post-baseline, after which time the divergence started to decrease so that at after approximately 8
SC
years the rates in the two subcohorts converged.
DISCUSSION
M AN U
Our study suggests that while the onset of pregnancy in RRMS patients results in a milder course (in terms of the relapse rate) and slower progression of the disease (in terms of the rate of conversion to irreversible disability) over several initial years (possibly through reducing the inflammatory component of RRMS), these improvements seem to be followed with a rebound of
TE D
relapse activity and rate of progression parameters. Thus, cumulatively over the 10-year time horizon, the course of illness in the pregnant and non-pregnant women appears to be similar. As for the prospects of transition to SPMS, our study suggests – albeit only weakly – that pregnancy
EP
may in fact increase the risk of this outcome throughout the 10-year period (possibly through promoting the neurodegenerative component of RRMS).
AC C
Several longitudinal studies have previously examined the effect of pregnancy on longterm disability, some of which suggested beneficial impact in RRMS patients [3,18,19]. However, the numbers of observations contributed by pregnant RRMS patients in those studies were typically so small that the findings could be attributable to random error; furthermore, the studies seem to have suffered from several important biases. Previous evidence on long-term influence of pregnancy on secondary progression and relapse rate is even sparser. We have
12
ACCEPTED MANUSCRIPT
identified only one longitudinal study that seems to have addressed it: Koch et al. [20] followed a cohort of 277 women with MS for an unspecified period of time, over which 49% of the patients developed SPMS. Based on a multiple logistic regression adjusting for several potential
RI PT
confounders, the estimated odds ratio for the association between parity and SPMS was 0.93 [20]. However, the results were very statistically imprecise and appear to have been subject to immortal-time bias, making the study virtually uninformative. As for longitudinal studies on
SC
long-term effects of pregnancy on the relapse rate, there has been, to our knowledge, only one
with the results being highly unreliable [21].
M AN U
such study, which however, was virtually uninformative, as it was based on a total of 37 subjects,
Our study examined the impact of pregnancy on the course of RRMS over a long period of time in a large cohort of women, with focus on three clinically- and patient-relevant outcomes. Still, because at issue is a young patient population, future studies should attempt to extend this
TE D
to even longer periods of observation. An accompanying challenge in non-experimental longitudinal studies of this kind is potential biases due to confounding and selective cohort attrition.
EP
Despite our reliance on modern longitudinal study-design methodology, our findings need to be interpreted with caution in the light of the following limitations. First, the potential
AC C
lack of comparability of the pregnant and non-pregnant cohorts in terms of prognostic factors enhances the possibility of non-causal explanations for the observed associations. While we attempted to minimize this possibility by accounting for a variety of prognostic factors, the success of this strategy relies on difficult-to-judge presumptions of valid documentation of the potential confounders, absence of unmeasured confounding, and correct model specification. Second, because of the long follow-up, the degree of cohort attrition was non-negligible. While
13
ACCEPTED MANUSCRIPT
we addressed this issue by applying the technique of inverse-probability of censoring weighting, so as to reconstruct the study experience had there been no attrition, this method also relies on strong presumptions, such as valid documentation of determinants of attrition, absence of
RI PT
unmeasured determinants of attrition, and correct model specification. Third, despite a large cohort size, the numbers of accrued events of progression to irreversible disability and transition to SPMS were relatively low, so that our results, especially those in the later segments of follow-
SC
up, were not precise enough to allow for more definitive inference. Fourth, we relied on medical records for data collection, so that the data on both the study outcomes and covariates were not
M AN U
completely accurate, although we believe the documentation errors were likely distributed equally between the two contrasted subcohorts and therefore conducive to bias towards the null in our results. In addition, t some data were missing. If, conditionally on the characteristics controlled for in our analyses, the patterns of missingness were different between the subcohorts,
TE D
it could have introduced some bias as well. Finally, our study addressed only the “total effect” of pregnancy, representing a net effect of several potential mechanisms, such as pregnancy-induced immunological changes, delivery, breastfeeding, and health behaviour. Again, there were not
EP
enough data to allow for mediation analysis and examination of the magnitudes of effect via different potential pathways.
AC C
In summary, while pregnancy likely ameliorates the short-term course of RRMS in terms of the rates of relapses and progression to irreversible disability, in the long term it appears to have no material impact on these outcomes, and might in fact accelerate the rate of transition to SPMS. The prognostic implications of pregnancy beyond the 10-year time horizon remain unclear. The current knowledge (even if it remains less than certain) on the impact of pregnancy on the course of RRMS should be discussed with patients contemplating pregnancy.
14
ACCEPTED MANUSCRIPT
ACKNOWLEDGEMENTS IK is a Canadian Institutes of Health Research New Investigator. AM was a recipient of a
AC C
EP
TE D
M AN U
SC
RI PT
summer studentship from the endMS Network of the Multiple Sclerosis Society of Canada.
15
ACCEPTED MANUSCRIPT
REFERENCES 1. Douglass LH, Jorgensen CL. Pregnancy and multiple sclerosis. Am J Obstet Gynegol 1948; 55:332-336.
RI PT
2. Runmarker B, Andersen O. Pregnancy is associated with a lower risk of onset and a better prognosis in multiple sclerosis. Brain 1995; 118:253-261.
3. Verdru P, Theys P, D’hooghe MB, Carton H. Pregnancy and multiple sclerosis: the
SC
influence on long term disability. Clin Neurol Neurosurg 1994; 96:38-41.
4. Birk K, Rudick R. Pregnancy and multiple sclerosis. Arch Neurol 1986; 43: 719-726.
M AN U
5. D’hooghe MB, Nagels G, Bissay T, De Keyser J. Modifiable factors influencing relapses and disability in multiple sclerosis. Mult Scler 2010; 16:773-785. 6. Finkelsztejn A, Brooks JBB, Paschoal FM, Fragoso YD. What can we really tell women with multiple sclerosis regarding pregnancy? A systematic review and meta-analysis of
TE D
the literature. BJOG 2011; 118:790-797.
7. Hernan MA, Alonso A, Logan R, et al. Observational studies analyzed like randomized experiments: an application to postmenopausal hormone therapy and coronary heart
EP
disease. Epidemiology 2008; 19:766-779.
8. Kurtzke JF. Further notes on disability evaluation in multiple sclerosis with scale
AC C
modifications. Neurology 1965; 11:686-694. 9. Cupples LA, D’Agostino RB, Anderson K, Kannel WB. Comparison of baseline and repeated measure covariate techniques in the Framingham Heart Study. Stat Med 1988; 7:205-218.
10. Rosenbaum PR, Rubin DB. The central role of the propensity score in observational studies for causal effects. Biometrika 1983; 70:41-55.
16
ACCEPTED MANUSCRIPT
11. Cole SR, Hernan MA. Constructing inverse probability weights for marginal structural models. Am J Epidemiol 2008; 168:656-664. 12. Westreich D, Cole SR. Invited commentary: Positivity in practice. Am J Epidemiol 2010;
RI PT
171:674-677.
13. D’Agostino RB, Lee ML, Belanger AJ, et al. Relation of pooled logistic regression to time-
9:1501-1515.
SC
dependent Cox regression analysis: the Framingham Heart Study. Stat Med 1990;
14. Robins JM, Hernan MA, Brumback B. Marginal structural models and causal inference
M AN U
in epidemiology. Epidemiology 2000; 11:550-560.
15. Pan W. Akaike's information criterion in generalized estimating equations. Biometrics 2001; 57:120-125.
16. Hernan MA. The hazards of hazard ratios. Epidemiology 2010; 21:13-15.
TE D
17. Miettinen OS. Theoretical epidemiology. New York: Wiley, 1985. 18. Stenager E, Stenager EN, Jensen K. Effect of pregnancy on the prognosis of multiple sclerosis. A 5-year follow up investigation. Acta Neurol Scand 1994; 90:305-308.
EP
19. D’Hooghe MB, Nagels G, Uitdehaag BMJ. Long-term effects of childbirth in MS. J Neurol Neurosurg Psychiatry 2010; 81:38-41.
AC C
20. Koch M, Uyttenboogart, Heersema D, Steen C, De Keyser J. Parity and secondary progression in multiple sclerosis. J Neurol Neurosurg Psychiatry 2009; 80:676-678. 21. Worthington J, Jones R, Crawford M, et al. Pregnancy and multiple sclerosis – a 3-year prospective study. J Neurol 1994; 241:228-233.
17
ACCEPTED MANUSCRIPT
Table 1. Distributions of selected characteristics of women with RRMS in the pregnant and nonpregnant sub-cohorts at baseline.
Pregnant (n=254) 30.7 (4.6)
Age at RRMS onset, mean (SD), years
23.4 (5.0)
Time since RRMS onset, means (SD), years
7.2 (4.5)
At least one relapse in the past three months, %
History of prior pregnancies, % Smoking, % History of use of oral contraceptives, %
Use of interferon beta, % Use of Copaxone, %
TE D
History of use of hormone therapy, %
31.5 (6.2) 25.5 (5.9) 6.0 (4.9)
10.2
22.4
0.98 (1.4)
0.72 (1.2)
M AN U
EDSS score, mean (SD)
SC
Age, mean (SD), years
Non-pregnant (n=423)
RI PT
Characteristic
63.8
78.3
22.4
27.7
55.1
39.2
11.4
14.4
11.0
7.8
4.3
2.8
RRMS = relapsing-remitting multiple sclerosis; SD = standard deviation; EDSS = Expanded
AC C
EP
Disability Status Scale.
18
ACCEPTED MANUSCRIPT
Table 2. The exponentiated estimated coefficients (95% CI) in the fitted pooled Poisson and logistic regression models for estimation of effect of pregnancy status on three study outcomes in women with RRMS.a
Progression to irreversible disability
Transition to SPMS
0.003 (0.002, 0.007)
RI PT
Relapse rate
Intercept
0.13 (0.11, 0.15)
0.005 (0.002, 0.008)
Pregnancy subcohort membership
0.67 (0.38, 0.65)
0.16 (0.04, 0.79)
Time (in years) since baseline
0.88 (0.85, 0.91)
0.93 (0.81, 1.08)
0.97 (0.86, 1.10)
Interaction term between pregnancy subcohort membership and time since baseline
1.09 (1.02, 1.17)
1.38 (1.06, 1.80)
1.14 (0.93, 1.39)
Propensity scoreb
0.60 (0.38, 0.94)
M AN U
Variable
1.40 (0.43, 4.58)
0.34 (0.08, 1.42)
a
SC
1.16 (0.36, 3.77)
AC C
EP
TE D
All models were weighted by the inverse of the probability of remaining uncensored by the end of a given 3-month interval. b The propensity score was based on the following baseline covariates: time since RRMS onset, time since RRMS onset squared, age, age squared, one relapse within the past three months, two or more relapses within the past three months, EDSS score, indicator of missing value of EDSS score, history of prior pregnancies, history of use of oral contraceptives, indicator of missing value of history of use of oral contraceptives, interaction term between EDSS score and time since RRMS onset, interaction term between history of prior pregnancies and age, interaction term between history of use of oral contraceptives and age, interaction term between history of use of oral contraceptives and EDSS score, interaction term between history of use of oral contraceptives and age squared, interaction term between history of prior pregnancies and EDSS score. CI = confidence interval; RRMS = relapsing-remitting multiple sclerosis; EDSS = Expanded Disability Status Scale.
19
ACCEPTED MANUSCRIPT
Figure 1. Adjusted relapse rates in women with relapsing-remitting multiple sclerosis as a
AC C
EP
TE D
M AN U
SC
RI PT
function of time since baseline, in the pregnant and non-pregnant subcohorts.
20
ACCEPTED MANUSCRIPT
Figure 2. Adjusted cumulative probabilities of progression to irreversible disability in women with relapsing-remitting multiple sclerosis as a function of time since baseline, in the pregnant
AC C
EP
TE D
M AN U
SC
RI PT
and non-pregnant subcohorts.
21
ACCEPTED MANUSCRIPT
Figure 3. Adjusted cumulative probabilities of transition to secondary progressive multiple sclerosis in women with relapsing-remitting multiple sclerosis as a function of time since
AC C
EP
TE D
M AN U
SC
RI PT
baseline, in the pregnant and non-pregnant subcohorts.
22
ACCEPTED MANUSCRIPT Supplementary Table 1. Results of the fitting of logistic regression model to estimate inputs into propensity scores. Standard error
Intercept
-5.6703
3.0827
Time (in years) since RRMS onset
0.4080
0.0661
Time (in years) since RRMS onset, squared
-0.0184
0.00394
Age (in years) at baseline
0.4371
0.1938
Age (in years) at baseline, squared
-0.00758
0.00310
One relapses within three months before baseline
-1.3638
0.2667
Two or more relapses within three months before baseline
RI PT
Point Estimate
SC
Variable
-2.3564
1.0811
-0.8119
0.2275
-2.7928
0.2558
0.8466
1.5433
-13.9618
5.2948
Interaction term EDSS score at baseline and time since RRMS onset
0.0556
0.0187
Interaction term between history of prior pregnancies at baseline and age at baseline
-0.0897
0.0523
Interaction term between history of use of oral contraceptives at baseline and age at baseline
0.8002
0.3364
Indicator of missing value of the history of use of oral contraceptives at baseline
0.8125
0.3358
Interaction term between history of the use of oral contraceptives at baseline and EDSS score at baseline
0.3949
0.1720
Interaction term between history of use of oral contraceptives at baseline and age squared at baseline
-0.0103
0.00527
Interaction term between history of prior pregnancies at baseline and EDSS score at baseline
-0.2746
0.1761
EDSS score at baseline History of prior pregnancies at baseline
M AN U
Indicator of missing value of EDSS score at baseline
AC C
EP
TE D
History of use of oral contraceptives at baseline
ACCEPTED MANUSCRIPT
Standard Error
Intercept
3.3204
0.3653
Time (in years) between RRMS onset and baseline
0.00486
0.0127
Age (in years) at baseline
0.00445
0.0107
History of use of oral contraceptives at baseline
-0.3032
0.1759
SC
Variable
-0.2609
0.1785
0.0449
0.1201
Indicator of missing value on smoking status at baseline
0.2619
0.2083
History of use of hormone therapy at baseline
0.1990
0.1507
Indicator of missing value on history of use of hormone therapy at baseline
-0.1647
0.1978
EDSS at baseline
0.1411
0.0676
Indicator of missing value on EDSS at baseline
0.4856
0.1638
History of prior pregnancies at baseline
1.3648
0.1602
Number of relapses within three months before baseline
-0.00465
0.1195
Use of interferon beta at baseline
-0.5140
0.2113
Use of Copaxone at baseline
0.2156
0.3204
Calendar year at baseline
-0.3387
0.0207
EDSS at baseline
-0.0163
0.0426
Indicator of missing value of EDSS at baseline
0.1004
0.1703
Number of relapses in the past three months
-0.1786
0.1863
Use of interferon beta in the past three months
0.4214
0.1336
Use of Copaxone in the past three months
0.3526
0.1644
Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
RI PT
Point Estimate
EP
Supplementary Table 2. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of relapse rates. The non-pregnant subcohort.
ACCEPTED MANUSCRIPT
Standard Error
Intercept
3.2600
0.5758
Time (in years) between RRMS onset and baseline
-0.0134
0.0169
0.000684
0.0174
-0.4816
0.2260
SC
Variable
-0.1997
0.2504
0.0748
0.1722
Indicator of missing value on smoking status at baseline
0.5090
0.3690
History of use of hormone therapy at baseline
0.2041
0.2153
Indicator of missing value on history of use of hormone therapy at baseline
0.1690
0.2929
EDSS at baseline
0.0933
0.0790
Indicator of missing value on EDSS at baseline
0.0244
0.2560
History of prior pregnancies at baseline
0.2290
0.1581
Number of relapses within three months before baseline
0.0180
0.2116
Use of interferon beta at baseline
-0.6838
0.2541
Use of Copaxone at baseline
-0.5641
0.4099
Calendar year at baseline
-0.0953
0.0253
EDSS at baseline
-0.0891
0.0742
Indicator of missing value of EDSS at baseline
0.0611
0.3250
Number of relapses in the past three months
-0.0138
0.2361
Use of interferon beta in the past three months
0.0693
0.2128
Use of Copaxone in the past three months
0.4164
0.2956
Age (in years) at baseline History of use of oral contraceptives at baseline Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
RI PT
Estimate
EP
Supplementary Table 3. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of relapse rates. The pregnant subcohort.
ACCEPTED MANUSCRIPT
Standard Error
Intercept
3.3405
0.3902
Time (in years) between RRMS onset and baseline
0.00356
0.0135
Age (in years) at baseline
0.000559
0.0113
-0.3485
0.1890
SC
Variable
-0.2449
0.1945
0.0460
0.1292
Indicator of missing value on smoking status at baseline
0.1562
0.2195
History of use of hormone therapy at baseline
0.1626
0.1630
Indicator of missing value on history of use of hormone therapy at baseline
-0.2201
EDSS at baseline
0.2216
0.0765
Indicator of missing value on EDSS at baseline
0.6334
0.1727
History of prior pregnancies at baseline
1.5232
0.1699
Number of relapses within three months before baseline
0.0485
0.1292
Use of interferon beta at baseline
-0.5534
0.2209
Use of Copaxone at baseline
0.2353
0.3341
Calendar year at baseline
-0.3491
0.0224
EDSS at baseline
0.0503
0.0478
Indicator of missing value of EDSS at baseline
0.2727
0.1861
Number of relapses in the past three months
-0.3299
0.1927
Use of interferon beta in the past three months
0.4350
0.1448
Use of Copaxone in the past three months
0.2289
0.1689
History of use of oral contraceptives at baseline Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
RI PT
Estimate
EP
Supplementary Table 4. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of transition to secondary progressive multiple sclerosis. The non-pregnant subcohort.
0.2139
ACCEPTED MANUSCRIPT Supplementary Table 5. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of transition to secondary progressive multiple sclerosis. The pregnant subcohort. Standard Error
4.1104
0.6273
Time (in years) between RRMS onset and baseline
-0.00511
0.0182
Age (in years) at baseline
-0.0224
0.0180
History of use of oral contraceptives at baseline
-0.8042
0.2702
-0.4499
SC
0.2933
0.0528
0.1850
Indicator of missing value on smoking status at baseline
0.4262
0.3781
History of use of hormone therapy at baseline
0.0934
0.2325
Indicator of missing value on history of use of hormone therapy at baseline
0.0748
EDSS at baseline
0.1101
0.0878
Indicator of missing value on EDSS at baseline
0.0663
0.2690
History of prior pregnancies at baseline
0.2555
0.1662
Number of relapses within three months before baseline
-0.0395
0.2257
EP
Intercept
Estimate
RI PT
Variable
Use of interferon beta at baseline
-0.7977
0.2661
Use of Copaxone at baseline
-0.6574
0.4199
Calendar year at baseline
-0.0867
0.0271
EDSS at baseline
0.0409
0.0827
Indicator of missing value of EDSS at baseline
0.3029
0.3520
Number of relapses in the past three months
0.0744
0.2723
Use of interferon beta in the past three months
0.1853
0.2287
Use of Copaxone in the past three months
0.5274
0.3162
Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
0.3003
ACCEPTED MANUSCRIPT
Standard Error
Intercept
3.5386
0.3514
Time (in years) between RRMS onset and baseline
0.0189
0.0125
Age (in years) at baseline
-0.00227
0.0103
History of use of oral contraceptives at baseline
-0.2225
0.1635
SC
Variable
-0.2086
0.1690
-0.0215
0.1107
Indicator of missing value on smoking status at baseline
0.2112
0.1995
History of use of hormone therapy at baseline
0.1985
0.1433
Indicator of missing value on history of use of hormone therapy at baseline
-0.1513
EDSS at baseline
0.0866
0.0650
Indicator of missing value on EDSS at baseline
0.2712
0.1570
History of prior pregnancies at baseline
1.0463
0.1509
Number of relapses within three months before baseline
-0.0914
0.1039
Use of interferon beta at baseline
-0.4379
0.2075
Use of Copaxone at baseline
0.2009
0.3159
Calendar year at baseline
-0.3023
0.0191
EDSS at baseline
0.0143
0.0389
Indicator of missing value of EDSS at baseline
0.2638
0.1571
Number of relapses in the past three months
-0.1745
0.1658
Use of interferon beta in the past three months
0.3836
0.1251
Use of Copaxone in the past three months
0.3308
0.1565
Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
RI PT
Estimate
EP
Supplementary Table 6. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of progression to irreversible disability. The non-pregnant subcohort.
0.1857
ACCEPTED MANUSCRIPT
Standard Error
Intercept
2.9221
0.5373
Time (in years) between RRMS onset and baseline
-0.0106
0.0160
Age (in years) at baseline
0.00759
0.0162
History of use of oral contraceptives at baseline
-0.4712
0.2199
SC
Variable
-0.1789
0.2440
0.1887
0.1601
Indicator of missing value on smoking status at baseline
0.2219
0.3238
History of use of hormone therapy at baseline
0.2809
0.2030
Indicator of missing value on history of use of hormone therapy at baseline
0.5412
EDSS at baseline
0.0621
0.0742
Indicator of missing value on EDSS at baseline
-0.0558
0.2452
History of prior pregnancies at baseline
0.2150
0.1390
Number of relapses within three months before baseline
0.0175
0.2048
Use of interferon beta at baseline
-0.4713
0.2265
Use of Copaxone at baseline
-0.5312
0.4051
Calendar year at baseline
-0.0899
0.0240
EDSS at baseline
-0.0676
0.0699
Indicator of missing value of EDSS at baseline
0.1910
0.3175
Number of relapses in the past three months
0.0964
0.2348
Use of interferon beta in the past three months
0.1078
0.1992
Use of Copaxone in the past three months
0.3833
0.2928
Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
RI PT
Estimate
EP
Supplementary Table 7. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of progression to irreversible disability. The pregnant subcohort.
0.2540
S1047-2797(14)00145-8
DOI:
10.1016/j.annepidem.2014.04.007
Reference:
AEP 7645
To appear in:
Annals of Epidemiology
Received Date: 21 December 2013 Revised Date:
21 March 2014
Accepted Date: 15 April 2014
Please cite this article as: Karp I, Manganas A, Sylvestre MP, Ho A, Roger E, Duquette P, Does pregnancy alter the long-term course of multiple sclerosis?, Annals of Epidemiology (2014), doi: 10.1016/j.annepidem.2014.04.007. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Title: Does pregnancy alter the long-term course of multiple sclerosis? Authors: Igor Karp, MD, MPH, PhD,1,2 Alexandra Manganas, BCom,3 Marie-Pierre Sylvestre,
RI PT
PhD,1,2 Annie Ho,3 Elaine Roger,4 Pierre Duquette, MD4,5
Affiliations: 1 Centre de recherche du Centre hospitalier de l'Université de Montréal
2
SC
(CRCHUM), Montréal, Québec, Canada
Département de médecine sociale et préventive, Université de Montréal, Montréal,
M AN U
Québec, Canada 3
Université de Montréal, Montréal, Québec, Canada
4
MS Clinic, Hôpital Notre-Dame du CHUM, Montréal, Québec, Canada
5
Faculté de Médecine (Neurology), Université de Montréal, Montréal, Québec,
TE D
Canada
Corresponding author: Igor Karp, Department of Social and Preventive Medicine, University
EP
of Montreal, 850 rue St Denis, #S02-366, Montreal, Quebec H2X 0A9. Tel: 514-890-8000, extension 15908 Fax: 514-412-7106
AC C
Email: [email protected]
Word count (main text only): 3141 Words in abstract: 225 Number of Tables: 2 Number of Supplementary Tables: 7 Number of Figures: 3
1
ACCEPTED MANUSCRIPT
Keywords: multiple sclerosis; pregnancy; prognosis; cohort; epidemiology; longitudinal studies List of abbreviations: RRMS, relapsing-remitting multiple sclerosis; SPMS, secondary progressive multiple sclerosis; EDSS, Expanded Disability Status Scale; RR, rate ratio; CI,
AC C
EP
TE D
M AN U
SC
RI PT
confidence interval;
2
ACCEPTED MANUSCRIPT
ABSTRACT Purpose: The purpose was to examine the impact of pregnancy on the rates of relapses,
RI PT
progression to irreversible disability, and transition to secondary progressive multiple sclerosis (SPMS) in patients with relapsing-remitting multiple sclerosis (RRMS).
Methods: We retrospectively followed two subcohorts of women with RRMS: pregnant
SC
(n=254) and non-pregnant (n=423). We obtained data on demographic, life-style, and clinical characteristics from patient records. Poisson and logistic regressions estimated the rate ratios
M AN U
(RR) associated with pregnancy as a function of time. Confounding was controlled by propensity-score adjustment, and post-baseline selection bias was controlled by inverse probability weighting.
Results: In the pregnant and non-pregnant subcohorts, respectively, 300 and 787 relapses, 15
TE D
and 27 transitions to SPMS, and 11 and 34 progressions to irreversible disability were documented. Adjusted RRs (95% confidence intervals) shortly after baseline were 0.67 (0.49; 0.92) for relapses, 0.16 (0.03; 0.79) for irreversible disability, and 1.25 (0.39; 3.96) for SPMS.
EP
The corresponding estimates at 5 and 10 years were, respectively, 1.04 (0.72; 1.52), 0.82 (0.36;
AC C
1.88), and 2.33 (1.03; 5.26) and 1.62 (0.84; 3.14), 4.14 (0.89; 19.22), and 4.33 (1.10; 16.99). Conclusions: Pregnancy likely ameliorates the short-term course of RRMS in terms of the rates of relapses and progression to irreversible disability. Over the long term it appears to have no material impact on these outcomes, and might in fact accelerate the rate of transition to SPMS.
3
ACCEPTED MANUSCRIPT
BACKGROUND Relapsing-remitting multiple sclerosis (RRMS) is an immune-mediated chronic inflammatory disease that is more common in women than in men. With a typical age of onset in the 3rd or 4th
RI PT
decades, the potential impact of pregnancy on RRMS course is of practical importance. While originally pregnancy was thought to worsen the course of the disease, [1] in the last several decades an alternative hypothesis was proposed, according to which pregnancy may actually
SC
have a favorable impact on RRMS [2,3].
Several epidemiological studies have attempted to address the competing hypotheses.
M AN U
However, the interpretation of, and the inference from, these studies has been difficult due to small sample sizes and various biases, of both individual studies (notably, confounding and selection bias) and their aggregate (i.e., publication bias). [4-6] Furthermore, most previous studies assessed only short-term consequences of pregnancy, usually in terms of a single
TE D
outcome. However, of even greater medical and scientific relevance would be knowledge of both the short and long-term impact of pregnancy on the triad of patient-relevant outcomes: relapse rate, transition to secondary progressive multiple sclerosis (SPMS) and progression to
EP
irreversible disability.
We set out to produce new evidence on the long-term impact of pregnancy on the course
AC C
of RRMS.
METHODS
Study population
We used data from medical records of 1,317 female RRMS patients followed at Notre-Dame Hospital Multiple Sclerosis Clinic in Montreal, Canada, from 1977 to 2010. Because at issue is a
4
ACCEPTED MANUSCRIPT
prognostic topic – namely, the prospective course of RRMS as a function of pregnancy status – data were structured as if pregnancy were a health intervention [7]. We retrospectively formed two subcohorts: the pregnant and the non-pregnant. Specifically, the non-pregnant subcohort was
RI PT
formed at the time of clinic entry (if the woman was not pregnant at that time), with this time point representing the subcohort’s baseline. The pregnant subcohort was formed among women in whom pregnancy occurred during follow-up, with the onset of pregnancy representing the
SC
pregnant subcohort’s baseline. Thus, each woman could contribute one observation to the nonpregnant subcohort and as many observations as the number of post-clinic-entry pregnancies to
M AN U
the pregnant subcohort. The inclusion criteria at baseline in both subcohorts were: (i) confirmed diagnosis of RRMS; (ii) age between 15 and 50 years, (iii) no menopause or history of hysterectomy; (iii) no prior transition to SPMS. Furthermore, in the analyses addressing the occurrence of progression to irreversible disability, the woman was not to have experienced that
TE D
outcome prior to baseline. Study outcomes
Three outcome measures were used: relapse rate (in analysis 1), transition to SMPS (in analysis
EP
2), and progression to irreversible disability (in analysis 3). The prevailing clinical practice in the Multiple Sclerosis clinic at the Hôpital Notre-Dame du CHUM, Montréal, the diagnosis of
AC C
SPMS is typically achieved when, in a patient with the Expanded Disability Status Scale (EDSS) score of 4 or higher, there is at least a 1-unit increase in that score over a period of at least 6 months without an intervening relapse. Progression to irreversible disability was operationalized as the first documented instance of the EDSS score being ≥4 [8]. Follow-up
5
ACCEPTED MANUSCRIPT
Each patient was followed from baseline until the occurrence of one of the following events, whichever came first: (i) a given study outcome (in the analyses addressing the occurrence of non-recurrent outcomes, i.e., in analyses 2 and 3); (ii) the end of study period (i.e., July 1st,
RI PT
2010); (iii) loss to follow-up, defined as no clinic visit for more than two years; (iv) transition to SPMS (in analyses 1 and 3); (v) the onset of post-baseline pregnancy; (vi) abortion. The latter five follow-up-terminating events thus represented a censoring mechanism.
SC
Statistical analysis
Data on follow-up for each woman were broken down in 3-month intervals and pooled across the
M AN U
subjects [9]. Due to the non-experimental nature of the study, two major threats to validity had to be addressed: confounding (due to potential non-comparability of the pregnant and non-pregnant women in terms of prognostic indicators) and prospective selection bias (due to potentially nonrandom censoring). Thus, before estimation of the associations between the study outcomes and
TE D
the pregnancy status, we first fitted two kinds of regression models: one, estimating the probability of membership in the pregnant subcohort as a function of its determinants (represented by the propensity score), [10] and the other, estimating the probability of censoring
EP
[11].
AC C
Estimation of propensity scores for control of confounding The dependent variable in the logistic model was a binary indicator of pregnancy subcohort membership (1 if yes, 0 if no). The potential confounders in the propensity score model were ascertained at baseline and included: time since RRMS onset, age, number of relapses in the past three months, history of prior pregnancies, EDSS score, smoking status, history of use of hormone therapy, history of use of oral contraceptives, use of interferon beta in the past three
6
ACCEPTED MANUSCRIPT
months, and use of Copaxone in the past three months. Furthermore, we added to these “maineffects” pairwise interaction terms for all the variables and quadratic terms for continuous variables. The propensity-score model was estimated according to the stepwise selection
RI PT
procedure applied to the above set of variable.
Before proceeding to the fitting of the association models of interest, we examined the distributions of the estimated propensity scores in the pregnant and non-pregnant cohorts, which
SC
revealed that they were only partially overlapping. However, inclusion of observations from nonoverlapping regions could potentially result in non-positivity bias [11]. To avoid this problem,
M AN U
we carried out “propensity score trimming” and restricted the final analytic set to observations with propensity scores between the 5th and 95th centile values [12].
Estimation of inverse-probability-of-censoring weights for control of selection bias
TE D
During follow-up, as the cohort underwent some attrition, we used the method of inverseprobability-of-censoring weighting to minimize the possibility of prospective selection bias [11]. In both subcohorts, a multiple pooled logistic regression was estimated [9,13]. The dependent
EP
variable was the event of not being censored by the end of the given 3-month interval. The independent variables included a set of variables measured at baseline (namely, time since
AC C
RRMS onset, age, number of relapses in the past three months, history of prior pregnancies, EDSS score, smoking status, history of use of hormone therapy, history of use of oral contraceptives, use of interferon beta in the past three months, and use of Copaxone in the past three months), supplemented with time-dependent variables: time since baseline, current EDSS score, number of relapses in the past three months, use of interferon beta in the past three months, and use of Copaxone in the past three months.
7
ACCEPTED MANUSCRIPT
Unstabilized weights were estimated by computing the reciprocal of the estimated probability of being uncensored in the given interval. To stabilize the weights, we estimated an analogous pooled logistic regression model with only time-invariant covariates so that stabilized
RI PT
weights were estimated by multiplying the unstabilized weights by the probability of being
Estimation of effect of pregnancy on the study outcomes
SC
uncensored, as estimated by this model [14].
Generalized linear models weighted by the inverse of the probability of being uncensored were
M AN U
estimated: pooled Poisson regression models in analysis 1 and pooled logistic regression models in analyses 2 and 3. All the models were estimated in the framework of generalized estimating equations to account for intra-individual correlation; the working correlation matrix was used. Since the probability of the non-recurrent study outcomes in a given 3-month interval is very
TE D
low, the pooled logistic regressions emulate Cox proportional-hazards regressions in estimating the RR [13].
The independent variables in all the models included an indicator of the pregnancy status,
EP
time since baseline, an interaction term between these two variables, and the propensity score. To account for potential non-linearity of the effect of time, we considered several alternative
AC C
approaches, namely: a quadratic time function as well as regression splines with one interior knot. Based on the Pan’s Information Criterion (QIC), the linear time function was retained [15]. Since interpretation of a time-function of RRs for non-recurrent outcomes can be challenging, [16] we estimated standardized cumulative incidence curves as a function of pregnancy status and time since baseline. Specifically, we estimated the predicted conditional survival probability for each 3-month interval as a function of codeterminants’ values at baseline
8
ACCEPTED MANUSCRIPT
and time since baseline for each woman under the scenario – actual or counterfactual – of membership in the pregnant subcohort. Sequential multiplication of these probabilities produced the estimates of cumulative survival for each woman up to each time point. The cumulative
RI PT
incidence estimate for each woman at each time point was obtained by deriving the complement of the corresponding cumulative survival estimate. Averaging the cumulative incidence estimates across all the observations produced the estimate of the cumulative incidence for pregnant
incidence for non-pregnant women was obtained similarly.
SC
women standardized to the whole study population. The standardized estimate of the cumulative
M AN U
Missing values of the independent variables were dealt with by the missing-value indicator method [17]. All analyses were performed using R 2.14.1 and SAS version 9.1 for Windows.
TE D
RESULTS
Of the 1,317 women in the source population, 1,026 met the non-pregnant subcohort membership criteria. During follow-up, there were documented 290 pregnancies in 183 women
EP
who met the pregnant subcohort membership criteria. After exclusion of observations where the estimated propensity score was below the 5th centile or above the 95th centile value, the number
AC C
of observations in the restricted analytic set was 254 (contributed by 165 women) in the pregnant subcohort and 423 in the non-pregnant subcohort. A total of 1,035 women-years were accrued in the pregnant subcohort and 2,655 women-years in the non-pregnant subcohort, with the median duration of follow-up being 2.9 and 6.9 years, respectively. Distributions of relevant characteristics of the subjects at baseline are summarized in Table 1. Compared with women in the non-pregnant subcohort, women in the pregnant
9
ACCEPTED MANUSCRIPT
subcohort were less likely to have been pregnant before (78% vs 64%), have a history of cigarette smoking (28% vs 22%) and hormone therapy (14% vs 11%), and to have had at least one relapse within three months before baseline(22% vs 10%), but more likely to have a history
RI PT
of use of oral contraceptives (39% vs 55%) and to have used disease-modifying drugs (DMDs) within three months before baseline(8% vs 11% for interferon beta and 3% vs 4% for Copaxone).
SC
During follow-up, in the pregnant and non-pregnant subcohorts, respectively, 300 and 787 relapses, 15 and 27 events of transition to SPMS, and 11 and 34 events of progression to
M AN U
irreversible disability were documented. The crude relapse rates, transition to SPMS, and progression to irreversible disability were, in the pregnant and non-pregnant subcohorts, 0.285 and 0.295, 0.014 and 0.010, and 0.012 and 0.016 events per person-year, respectively. Results of regression modelling for estimation of propensity scores are shown in
TE D
Supplementary Table 1. The model had a high discriminating ability and the distributions of the potential confounders between the subcohorts within categories of the propensity score quartiles were well-balanced.
EP
Results of regression modelling for estimation of censoring weights are shown in Supplementary Tables 2-7. All the models were characterized by a moderate discriminating
AC C
ability, with narrowly distributed stabilized weights, suggesting that censoring was not heavily influenced by the characteristics at issue. Results of regression models for estimation of effects of pregnancy are shown in Table 2. The estimate of the adjusted relapse RRs (95% CI) for the pregnant vs non-pregnant shortly after baseline was 0.67 (0.38, 0.65), implying that at that time, the adjusted relapse rate in the pregnant subcohort was 50% lower than in the non-pregnant. The exponentiated estimated product-term
10
ACCEPTED MANUSCRIPT
coefficient for the subcohort membership status and time since baseline was 1.09 (1.02, 1.17), implying that with each additional year of follow-up, the adjusted relapse RR increased by this factor. Consequently, the estimates of the adjusted relapse RRs at 5 and 10 years post-baseline
RI PT
were 1.04 (0.72; 1.52) and 1.62 (0.84; 3.14), respectively. Figure 1 shows that the relapse rate decreased over time in both subcohorts and that while the relapse rate in the pregnant subcohort was lower in the first several years, the difference gradually diminished to the point of
SC
convergence around 8 years after baseline.
With the transition to SPMS as the outcome, the estimated adjusted RR (95% CI) for the
M AN U
pregnant vs non-pregnant shortly after the baseline was 1.16 (0.36, 3.77) and the exponentiated estimated product-term between the pregnancy subcohort status and time since baseline was 1.14 (0.93, 1.39), implying that there was virtually no evidence that rates of transition to SPMS shortly after the baseline are different and only weak evidence that the ratio between these rates
TE D
increases over time. Accordingly, the estimated adjusted RRs (95% CIs) at 5 and 10 years postbaseline were 2.33 (1.03; 5.26) and 4.33 (1.10; 16.99), respectively. Figure 2 shows that the divergence in cumulative incidence of transition to SPMS emerged soon after the baseline and
EP
continued throughout the follow-up, with the estimated 10-year cumulative incidence rates reaching 17% and 8% in the pregnant and non-pregnant subcohorts, respectively.
AC C
With conversion to irreversible disability as the outcome, the estimate of the adjusted RR (95% CI) for the pregnant vs non-pregnant shortly was 0.16 (0.04, 0.79) and the exponentiated estimated product-term between pregnancy subcohort status and time since baseline was 1.38 (1.06, 1.80), implying that there was moderately strong evidence that the rate of conversion to irreversible disability is lower in the pregnant than in the non-pregnant shortly after the baseline and that, over time, the rate in the pregnant decreases at a slower rate than in the non-pregnant.
11
ACCEPTED MANUSCRIPT
Accordingly, the estimated adjusted RRs (95% CIs) at 5 and 10 years post-baseline were 0.82 (0.36; 1.88), and 4.14 (0.89; 19.22), respectively. Figure 3 shows that the cumulative incidence rate of this outcome in the pregnant subcohort was lower than in the non-pregnant subcohort for
RI PT
the most of the follow-up, with the difference in cumulative incidence peaking at about 5 years post-baseline, after which time the divergence started to decrease so that at after approximately 8
SC
years the rates in the two subcohorts converged.
DISCUSSION
M AN U
Our study suggests that while the onset of pregnancy in RRMS patients results in a milder course (in terms of the relapse rate) and slower progression of the disease (in terms of the rate of conversion to irreversible disability) over several initial years (possibly through reducing the inflammatory component of RRMS), these improvements seem to be followed with a rebound of
TE D
relapse activity and rate of progression parameters. Thus, cumulatively over the 10-year time horizon, the course of illness in the pregnant and non-pregnant women appears to be similar. As for the prospects of transition to SPMS, our study suggests – albeit only weakly – that pregnancy
EP
may in fact increase the risk of this outcome throughout the 10-year period (possibly through promoting the neurodegenerative component of RRMS).
AC C
Several longitudinal studies have previously examined the effect of pregnancy on longterm disability, some of which suggested beneficial impact in RRMS patients [3,18,19]. However, the numbers of observations contributed by pregnant RRMS patients in those studies were typically so small that the findings could be attributable to random error; furthermore, the studies seem to have suffered from several important biases. Previous evidence on long-term influence of pregnancy on secondary progression and relapse rate is even sparser. We have
12
ACCEPTED MANUSCRIPT
identified only one longitudinal study that seems to have addressed it: Koch et al. [20] followed a cohort of 277 women with MS for an unspecified period of time, over which 49% of the patients developed SPMS. Based on a multiple logistic regression adjusting for several potential
RI PT
confounders, the estimated odds ratio for the association between parity and SPMS was 0.93 [20]. However, the results were very statistically imprecise and appear to have been subject to immortal-time bias, making the study virtually uninformative. As for longitudinal studies on
SC
long-term effects of pregnancy on the relapse rate, there has been, to our knowledge, only one
with the results being highly unreliable [21].
M AN U
such study, which however, was virtually uninformative, as it was based on a total of 37 subjects,
Our study examined the impact of pregnancy on the course of RRMS over a long period of time in a large cohort of women, with focus on three clinically- and patient-relevant outcomes. Still, because at issue is a young patient population, future studies should attempt to extend this
TE D
to even longer periods of observation. An accompanying challenge in non-experimental longitudinal studies of this kind is potential biases due to confounding and selective cohort attrition.
EP
Despite our reliance on modern longitudinal study-design methodology, our findings need to be interpreted with caution in the light of the following limitations. First, the potential
AC C
lack of comparability of the pregnant and non-pregnant cohorts in terms of prognostic factors enhances the possibility of non-causal explanations for the observed associations. While we attempted to minimize this possibility by accounting for a variety of prognostic factors, the success of this strategy relies on difficult-to-judge presumptions of valid documentation of the potential confounders, absence of unmeasured confounding, and correct model specification. Second, because of the long follow-up, the degree of cohort attrition was non-negligible. While
13
ACCEPTED MANUSCRIPT
we addressed this issue by applying the technique of inverse-probability of censoring weighting, so as to reconstruct the study experience had there been no attrition, this method also relies on strong presumptions, such as valid documentation of determinants of attrition, absence of
RI PT
unmeasured determinants of attrition, and correct model specification. Third, despite a large cohort size, the numbers of accrued events of progression to irreversible disability and transition to SPMS were relatively low, so that our results, especially those in the later segments of follow-
SC
up, were not precise enough to allow for more definitive inference. Fourth, we relied on medical records for data collection, so that the data on both the study outcomes and covariates were not
M AN U
completely accurate, although we believe the documentation errors were likely distributed equally between the two contrasted subcohorts and therefore conducive to bias towards the null in our results. In addition, t some data were missing. If, conditionally on the characteristics controlled for in our analyses, the patterns of missingness were different between the subcohorts,
TE D
it could have introduced some bias as well. Finally, our study addressed only the “total effect” of pregnancy, representing a net effect of several potential mechanisms, such as pregnancy-induced immunological changes, delivery, breastfeeding, and health behaviour. Again, there were not
EP
enough data to allow for mediation analysis and examination of the magnitudes of effect via different potential pathways.
AC C
In summary, while pregnancy likely ameliorates the short-term course of RRMS in terms of the rates of relapses and progression to irreversible disability, in the long term it appears to have no material impact on these outcomes, and might in fact accelerate the rate of transition to SPMS. The prognostic implications of pregnancy beyond the 10-year time horizon remain unclear. The current knowledge (even if it remains less than certain) on the impact of pregnancy on the course of RRMS should be discussed with patients contemplating pregnancy.
14
ACCEPTED MANUSCRIPT
ACKNOWLEDGEMENTS IK is a Canadian Institutes of Health Research New Investigator. AM was a recipient of a
AC C
EP
TE D
M AN U
SC
RI PT
summer studentship from the endMS Network of the Multiple Sclerosis Society of Canada.
15
ACCEPTED MANUSCRIPT
REFERENCES 1. Douglass LH, Jorgensen CL. Pregnancy and multiple sclerosis. Am J Obstet Gynegol 1948; 55:332-336.
RI PT
2. Runmarker B, Andersen O. Pregnancy is associated with a lower risk of onset and a better prognosis in multiple sclerosis. Brain 1995; 118:253-261.
3. Verdru P, Theys P, D’hooghe MB, Carton H. Pregnancy and multiple sclerosis: the
SC
influence on long term disability. Clin Neurol Neurosurg 1994; 96:38-41.
4. Birk K, Rudick R. Pregnancy and multiple sclerosis. Arch Neurol 1986; 43: 719-726.
M AN U
5. D’hooghe MB, Nagels G, Bissay T, De Keyser J. Modifiable factors influencing relapses and disability in multiple sclerosis. Mult Scler 2010; 16:773-785. 6. Finkelsztejn A, Brooks JBB, Paschoal FM, Fragoso YD. What can we really tell women with multiple sclerosis regarding pregnancy? A systematic review and meta-analysis of
TE D
the literature. BJOG 2011; 118:790-797.
7. Hernan MA, Alonso A, Logan R, et al. Observational studies analyzed like randomized experiments: an application to postmenopausal hormone therapy and coronary heart
EP
disease. Epidemiology 2008; 19:766-779.
8. Kurtzke JF. Further notes on disability evaluation in multiple sclerosis with scale
AC C
modifications. Neurology 1965; 11:686-694. 9. Cupples LA, D’Agostino RB, Anderson K, Kannel WB. Comparison of baseline and repeated measure covariate techniques in the Framingham Heart Study. Stat Med 1988; 7:205-218.
10. Rosenbaum PR, Rubin DB. The central role of the propensity score in observational studies for causal effects. Biometrika 1983; 70:41-55.
16
ACCEPTED MANUSCRIPT
11. Cole SR, Hernan MA. Constructing inverse probability weights for marginal structural models. Am J Epidemiol 2008; 168:656-664. 12. Westreich D, Cole SR. Invited commentary: Positivity in practice. Am J Epidemiol 2010;
RI PT
171:674-677.
13. D’Agostino RB, Lee ML, Belanger AJ, et al. Relation of pooled logistic regression to time-
9:1501-1515.
SC
dependent Cox regression analysis: the Framingham Heart Study. Stat Med 1990;
14. Robins JM, Hernan MA, Brumback B. Marginal structural models and causal inference
M AN U
in epidemiology. Epidemiology 2000; 11:550-560.
15. Pan W. Akaike's information criterion in generalized estimating equations. Biometrics 2001; 57:120-125.
16. Hernan MA. The hazards of hazard ratios. Epidemiology 2010; 21:13-15.
TE D
17. Miettinen OS. Theoretical epidemiology. New York: Wiley, 1985. 18. Stenager E, Stenager EN, Jensen K. Effect of pregnancy on the prognosis of multiple sclerosis. A 5-year follow up investigation. Acta Neurol Scand 1994; 90:305-308.
EP
19. D’Hooghe MB, Nagels G, Uitdehaag BMJ. Long-term effects of childbirth in MS. J Neurol Neurosurg Psychiatry 2010; 81:38-41.
AC C
20. Koch M, Uyttenboogart, Heersema D, Steen C, De Keyser J. Parity and secondary progression in multiple sclerosis. J Neurol Neurosurg Psychiatry 2009; 80:676-678. 21. Worthington J, Jones R, Crawford M, et al. Pregnancy and multiple sclerosis – a 3-year prospective study. J Neurol 1994; 241:228-233.
17
ACCEPTED MANUSCRIPT
Table 1. Distributions of selected characteristics of women with RRMS in the pregnant and nonpregnant sub-cohorts at baseline.
Pregnant (n=254) 30.7 (4.6)
Age at RRMS onset, mean (SD), years
23.4 (5.0)
Time since RRMS onset, means (SD), years
7.2 (4.5)
At least one relapse in the past three months, %
History of prior pregnancies, % Smoking, % History of use of oral contraceptives, %
Use of interferon beta, % Use of Copaxone, %
TE D
History of use of hormone therapy, %
31.5 (6.2) 25.5 (5.9) 6.0 (4.9)
10.2
22.4
0.98 (1.4)
0.72 (1.2)
M AN U
EDSS score, mean (SD)
SC
Age, mean (SD), years
Non-pregnant (n=423)
RI PT
Characteristic
63.8
78.3
22.4
27.7
55.1
39.2
11.4
14.4
11.0
7.8
4.3
2.8
RRMS = relapsing-remitting multiple sclerosis; SD = standard deviation; EDSS = Expanded
AC C
EP
Disability Status Scale.
18
ACCEPTED MANUSCRIPT
Table 2. The exponentiated estimated coefficients (95% CI) in the fitted pooled Poisson and logistic regression models for estimation of effect of pregnancy status on three study outcomes in women with RRMS.a
Progression to irreversible disability
Transition to SPMS
0.003 (0.002, 0.007)
RI PT
Relapse rate
Intercept
0.13 (0.11, 0.15)
0.005 (0.002, 0.008)
Pregnancy subcohort membership
0.67 (0.38, 0.65)
0.16 (0.04, 0.79)
Time (in years) since baseline
0.88 (0.85, 0.91)
0.93 (0.81, 1.08)
0.97 (0.86, 1.10)
Interaction term between pregnancy subcohort membership and time since baseline
1.09 (1.02, 1.17)
1.38 (1.06, 1.80)
1.14 (0.93, 1.39)
Propensity scoreb
0.60 (0.38, 0.94)
M AN U
Variable
1.40 (0.43, 4.58)
0.34 (0.08, 1.42)
a
SC
1.16 (0.36, 3.77)
AC C
EP
TE D
All models were weighted by the inverse of the probability of remaining uncensored by the end of a given 3-month interval. b The propensity score was based on the following baseline covariates: time since RRMS onset, time since RRMS onset squared, age, age squared, one relapse within the past three months, two or more relapses within the past three months, EDSS score, indicator of missing value of EDSS score, history of prior pregnancies, history of use of oral contraceptives, indicator of missing value of history of use of oral contraceptives, interaction term between EDSS score and time since RRMS onset, interaction term between history of prior pregnancies and age, interaction term between history of use of oral contraceptives and age, interaction term between history of use of oral contraceptives and EDSS score, interaction term between history of use of oral contraceptives and age squared, interaction term between history of prior pregnancies and EDSS score. CI = confidence interval; RRMS = relapsing-remitting multiple sclerosis; EDSS = Expanded Disability Status Scale.
19
ACCEPTED MANUSCRIPT
Figure 1. Adjusted relapse rates in women with relapsing-remitting multiple sclerosis as a
AC C
EP
TE D
M AN U
SC
RI PT
function of time since baseline, in the pregnant and non-pregnant subcohorts.
20
ACCEPTED MANUSCRIPT
Figure 2. Adjusted cumulative probabilities of progression to irreversible disability in women with relapsing-remitting multiple sclerosis as a function of time since baseline, in the pregnant
AC C
EP
TE D
M AN U
SC
RI PT
and non-pregnant subcohorts.
21
ACCEPTED MANUSCRIPT
Figure 3. Adjusted cumulative probabilities of transition to secondary progressive multiple sclerosis in women with relapsing-remitting multiple sclerosis as a function of time since
AC C
EP
TE D
M AN U
SC
RI PT
baseline, in the pregnant and non-pregnant subcohorts.
22
ACCEPTED MANUSCRIPT Supplementary Table 1. Results of the fitting of logistic regression model to estimate inputs into propensity scores. Standard error
Intercept
-5.6703
3.0827
Time (in years) since RRMS onset
0.4080
0.0661
Time (in years) since RRMS onset, squared
-0.0184
0.00394
Age (in years) at baseline
0.4371
0.1938
Age (in years) at baseline, squared
-0.00758
0.00310
One relapses within three months before baseline
-1.3638
0.2667
Two or more relapses within three months before baseline
RI PT
Point Estimate
SC
Variable
-2.3564
1.0811
-0.8119
0.2275
-2.7928
0.2558
0.8466
1.5433
-13.9618
5.2948
Interaction term EDSS score at baseline and time since RRMS onset
0.0556
0.0187
Interaction term between history of prior pregnancies at baseline and age at baseline
-0.0897
0.0523
Interaction term between history of use of oral contraceptives at baseline and age at baseline
0.8002
0.3364
Indicator of missing value of the history of use of oral contraceptives at baseline
0.8125
0.3358
Interaction term between history of the use of oral contraceptives at baseline and EDSS score at baseline
0.3949
0.1720
Interaction term between history of use of oral contraceptives at baseline and age squared at baseline
-0.0103
0.00527
Interaction term between history of prior pregnancies at baseline and EDSS score at baseline
-0.2746
0.1761
EDSS score at baseline History of prior pregnancies at baseline
M AN U
Indicator of missing value of EDSS score at baseline
AC C
EP
TE D
History of use of oral contraceptives at baseline
ACCEPTED MANUSCRIPT
Standard Error
Intercept
3.3204
0.3653
Time (in years) between RRMS onset and baseline
0.00486
0.0127
Age (in years) at baseline
0.00445
0.0107
History of use of oral contraceptives at baseline
-0.3032
0.1759
SC
Variable
-0.2609
0.1785
0.0449
0.1201
Indicator of missing value on smoking status at baseline
0.2619
0.2083
History of use of hormone therapy at baseline
0.1990
0.1507
Indicator of missing value on history of use of hormone therapy at baseline
-0.1647
0.1978
EDSS at baseline
0.1411
0.0676
Indicator of missing value on EDSS at baseline
0.4856
0.1638
History of prior pregnancies at baseline
1.3648
0.1602
Number of relapses within three months before baseline
-0.00465
0.1195
Use of interferon beta at baseline
-0.5140
0.2113
Use of Copaxone at baseline
0.2156
0.3204
Calendar year at baseline
-0.3387
0.0207
EDSS at baseline
-0.0163
0.0426
Indicator of missing value of EDSS at baseline
0.1004
0.1703
Number of relapses in the past three months
-0.1786
0.1863
Use of interferon beta in the past three months
0.4214
0.1336
Use of Copaxone in the past three months
0.3526
0.1644
Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
RI PT
Point Estimate
EP
Supplementary Table 2. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of relapse rates. The non-pregnant subcohort.
ACCEPTED MANUSCRIPT
Standard Error
Intercept
3.2600
0.5758
Time (in years) between RRMS onset and baseline
-0.0134
0.0169
0.000684
0.0174
-0.4816
0.2260
SC
Variable
-0.1997
0.2504
0.0748
0.1722
Indicator of missing value on smoking status at baseline
0.5090
0.3690
History of use of hormone therapy at baseline
0.2041
0.2153
Indicator of missing value on history of use of hormone therapy at baseline
0.1690
0.2929
EDSS at baseline
0.0933
0.0790
Indicator of missing value on EDSS at baseline
0.0244
0.2560
History of prior pregnancies at baseline
0.2290
0.1581
Number of relapses within three months before baseline
0.0180
0.2116
Use of interferon beta at baseline
-0.6838
0.2541
Use of Copaxone at baseline
-0.5641
0.4099
Calendar year at baseline
-0.0953
0.0253
EDSS at baseline
-0.0891
0.0742
Indicator of missing value of EDSS at baseline
0.0611
0.3250
Number of relapses in the past three months
-0.0138
0.2361
Use of interferon beta in the past three months
0.0693
0.2128
Use of Copaxone in the past three months
0.4164
0.2956
Age (in years) at baseline History of use of oral contraceptives at baseline Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
RI PT
Estimate
EP
Supplementary Table 3. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of relapse rates. The pregnant subcohort.
ACCEPTED MANUSCRIPT
Standard Error
Intercept
3.3405
0.3902
Time (in years) between RRMS onset and baseline
0.00356
0.0135
Age (in years) at baseline
0.000559
0.0113
-0.3485
0.1890
SC
Variable
-0.2449
0.1945
0.0460
0.1292
Indicator of missing value on smoking status at baseline
0.1562
0.2195
History of use of hormone therapy at baseline
0.1626
0.1630
Indicator of missing value on history of use of hormone therapy at baseline
-0.2201
EDSS at baseline
0.2216
0.0765
Indicator of missing value on EDSS at baseline
0.6334
0.1727
History of prior pregnancies at baseline
1.5232
0.1699
Number of relapses within three months before baseline
0.0485
0.1292
Use of interferon beta at baseline
-0.5534
0.2209
Use of Copaxone at baseline
0.2353
0.3341
Calendar year at baseline
-0.3491
0.0224
EDSS at baseline
0.0503
0.0478
Indicator of missing value of EDSS at baseline
0.2727
0.1861
Number of relapses in the past three months
-0.3299
0.1927
Use of interferon beta in the past three months
0.4350
0.1448
Use of Copaxone in the past three months
0.2289
0.1689
History of use of oral contraceptives at baseline Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
RI PT
Estimate
EP
Supplementary Table 4. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of transition to secondary progressive multiple sclerosis. The non-pregnant subcohort.
0.2139
ACCEPTED MANUSCRIPT Supplementary Table 5. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of transition to secondary progressive multiple sclerosis. The pregnant subcohort. Standard Error
4.1104
0.6273
Time (in years) between RRMS onset and baseline
-0.00511
0.0182
Age (in years) at baseline
-0.0224
0.0180
History of use of oral contraceptives at baseline
-0.8042
0.2702
-0.4499
SC
0.2933
0.0528
0.1850
Indicator of missing value on smoking status at baseline
0.4262
0.3781
History of use of hormone therapy at baseline
0.0934
0.2325
Indicator of missing value on history of use of hormone therapy at baseline
0.0748
EDSS at baseline
0.1101
0.0878
Indicator of missing value on EDSS at baseline
0.0663
0.2690
History of prior pregnancies at baseline
0.2555
0.1662
Number of relapses within three months before baseline
-0.0395
0.2257
EP
Intercept
Estimate
RI PT
Variable
Use of interferon beta at baseline
-0.7977
0.2661
Use of Copaxone at baseline
-0.6574
0.4199
Calendar year at baseline
-0.0867
0.0271
EDSS at baseline
0.0409
0.0827
Indicator of missing value of EDSS at baseline
0.3029
0.3520
Number of relapses in the past three months
0.0744
0.2723
Use of interferon beta in the past three months
0.1853
0.2287
Use of Copaxone in the past three months
0.5274
0.3162
Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
0.3003
ACCEPTED MANUSCRIPT
Standard Error
Intercept
3.5386
0.3514
Time (in years) between RRMS onset and baseline
0.0189
0.0125
Age (in years) at baseline
-0.00227
0.0103
History of use of oral contraceptives at baseline
-0.2225
0.1635
SC
Variable
-0.2086
0.1690
-0.0215
0.1107
Indicator of missing value on smoking status at baseline
0.2112
0.1995
History of use of hormone therapy at baseline
0.1985
0.1433
Indicator of missing value on history of use of hormone therapy at baseline
-0.1513
EDSS at baseline
0.0866
0.0650
Indicator of missing value on EDSS at baseline
0.2712
0.1570
History of prior pregnancies at baseline
1.0463
0.1509
Number of relapses within three months before baseline
-0.0914
0.1039
Use of interferon beta at baseline
-0.4379
0.2075
Use of Copaxone at baseline
0.2009
0.3159
Calendar year at baseline
-0.3023
0.0191
EDSS at baseline
0.0143
0.0389
Indicator of missing value of EDSS at baseline
0.2638
0.1571
Number of relapses in the past three months
-0.1745
0.1658
Use of interferon beta in the past three months
0.3836
0.1251
Use of Copaxone in the past three months
0.3308
0.1565
Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
RI PT
Estimate
EP
Supplementary Table 6. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of progression to irreversible disability. The non-pregnant subcohort.
0.1857
ACCEPTED MANUSCRIPT
Standard Error
Intercept
2.9221
0.5373
Time (in years) between RRMS onset and baseline
-0.0106
0.0160
Age (in years) at baseline
0.00759
0.0162
History of use of oral contraceptives at baseline
-0.4712
0.2199
SC
Variable
-0.1789
0.2440
0.1887
0.1601
Indicator of missing value on smoking status at baseline
0.2219
0.3238
History of use of hormone therapy at baseline
0.2809
0.2030
Indicator of missing value on history of use of hormone therapy at baseline
0.5412
EDSS at baseline
0.0621
0.0742
Indicator of missing value on EDSS at baseline
-0.0558
0.2452
History of prior pregnancies at baseline
0.2150
0.1390
Number of relapses within three months before baseline
0.0175
0.2048
Use of interferon beta at baseline
-0.4713
0.2265
Use of Copaxone at baseline
-0.5312
0.4051
Calendar year at baseline
-0.0899
0.0240
EDSS at baseline
-0.0676
0.0699
Indicator of missing value of EDSS at baseline
0.1910
0.3175
Number of relapses in the past three months
0.0964
0.2348
Use of interferon beta in the past three months
0.1078
0.1992
Use of Copaxone in the past three months
0.3833
0.2928
Indicator of missing data on history of use of oral contraceptives at baseline
AC C
TE D
M AN U
Smoking status at baseline
RI PT
Estimate
EP
Supplementary Table 7. Results of the fitting of the pooled logistic regression model to estimate the denominator inputs into inverse-probability-of-censoring weights in the analysis of progression to irreversible disability. The pregnant subcohort.
0.2540
