Search

Patients with polycythemia vera (PV) are at significant risk of thromboembolic events (TE). The PV-AIM study used the Optum® de-identified Electronic Health Record dataset and machine learning to identify markers of TE in a real-world population. Data for 82,960 patients with PV were extracted: 3852 patients were treated with hydroxyurea (HU) only, while 130 patients were treated with HU and then changed to ruxolitinib (HU-ruxolitinib). For HU-alone patients, the annualized incidence rates (IR; per 100 patients) decreased from 8.7 (before HU) to 5.6 (during HU) but increased markedly to 10.5 (continuing HU). Whereas for HU-ruxolitinib patients, the IR decreased from 10.8 (before HU) to 8.4 (during HU) and was maintained at 8.3 (after switching to ruxolitinib). To better understand markers associated with TE risk, we built a machine-learning model for HU-alone patients and validated it using an independent dataset. The model identified lymphocyte percentage (LYP), neutrophil percentage (NEP), and red cell distribution width (RDW) as key markers of TE risk, and optimal thresholds for these markers were established, from which a decision tree was derived. Using these widely used laboratory markers, the decision tree could be used to identify patients at high risk for TE, facilitate treatment decisions, and optimize patient management.

Aims: We previously reported results of the first machine learning study to identify the novel biomarkers associated with crude incidence of TEs in PV patients (pts) treated with hydroxyurea (HU) (Verstovsek et al, Blood 2019). In the current study, we have expanded the database to perform an in-depth machine learning analysis of biomarkers predicting the annual standardized incidence rate (IR) of TEs in PV pts treated with HU. In addition, we examined the IR of TEs in HU-treated pts and pts treated first with HU, then switched to ruxolitinib (HU-RUX). Methods: The study is part of PV-AIM (polycythemia vera advanced integrated models for prevention of TEs). The US OPTUM database includes the electronic medical records from 105 million pts (2007-2020), including 82 960 PV pts with a median record length of 8.4 years. TEs were assessed (A) before the treatment (Tx) initiation in both groups; (B) while on HU (median: 29 months) and (C) while continuing HU or switching to RUX (median, 17 months). A total of 3852 HU-alone pts and 130 HU-RUX pts passed inclusion/exclusion criteria. To avoid selection bias, only HU-alone pts treated prior to RUX market launch were included (n=704). Cohorts were then matched by selecting the nearest cases from the HU-alone groups based on the propensity scores calculated from the total Tx period, gender, race, age at index, and geographical division (n=130 pts in both HU-alone and HU-RUX). TEs were identified from a restrictive list of ICD-CM diagnosis codes defined in the RESPONSE study (Kiladjian et al, Lancet Haematol 2020). Annualized IR of TEs were then calculated per 100 pts for each of the above Tx periods for each cohort. A random survival forest (RSF) model was then constructed for HU-alone pts, with at least 6 months of HU Tx and 18 months of follow-up, to predict the risk of TEs 6 to 18 months after the first HU Tx. Pts with at least 1 lab test and 1 observation obtained at 3 to 6 months post index were included in the model (n=1012). The features also included demographics, history of TEs, phlebotomy and anti-coagulant use. The performance was assessed on a 70:30 train:test split using Receiver Operating Characteristic-Area Under the Curve (ROC-AUC). RSF variable importance was then used to identify the variables with the largest impact on prediction of TEs. The feature interactions for the top 10 features were mapped and assessed in terms of risk of TEs via log-rank using a brute-force combinatorial approach. Risk was assessed in the context of total population and in pts with/without a history of TEs. Optimal interactions were presented as interpretable 4-group Kaplan-Meier plots. Lastly, the feature interactions were scored for synergy in the context of risk of TEs based on the difference in expected and observed pairwise probabilities. Results: The RSF model achieved an accuracy of > 0.80 AUC for prediction of TEs 6 to12 months after the first HU Tx. RSF variable importance revealed that the history of TEs had the largest effect on prediction (>2-fold) relative to other features. The top observational and lab variables in terms of feature importance were body mass index and neutrophil percentage (NEP), respectively, followed by the previously observed lymphocyte percentage (LYP) and RDW. The largest synergistic pairwise interactions (Table 1) were NEP + RDW (Figure 1a), followed by LYP + RDW (Figure 1b), which occurred in pts without history of TEs. The pre-Tx baseline incidence of TEs in the HU-alone and HU-RUX groups were 8.7 and 10.8, respectively. The rates of TEs decreased by 36% and 22% to 5.6 and 8.4, respectively during the HU Tx. Following the switch to RUX, the rate of TEs was maintained in the HU-RUX group at 8.3 (−1%) but increased to 10.53 (+89%) in the HU-alone group during the same period of 17 months (Figure 2). Conclusion: An in-depth machine learning analysis to identify the markers that predict TEs in HU-treated PV pts revealed interaction of key hematological parameters with RDW. This finding could help recognize the pts at high risk of TE and possibly prevent a major cause of morbidity and mortality. The results reconfirm the role of RDW, as well as NEP/LYM ratios in predicting TE occurrence, routine lab parameters in the clinic, but not usually taken into account. Finally, we observed a renewed increase in TEs in HU-treated pts after an initial decrease, which was not the case in pts switching to RUX in the same period. We aim to present similar biomarker analyses on RUX pts in the final presentation. Disclosures Verstovsek: Promedior: Research Funding; NS Pharma: Research Funding; PharmaEssentia: Research Funding; Novartis: Consultancy, Research Funding; Sierra Oncology: Consultancy, Research Funding; AstraZeneca: Research Funding; Blueprint Medicines Corp: Research Funding; Genentech: Research Funding; CTI Biopharma Corp: Research Funding; ItalPharma: Research Funding; Incyte Corporation: Consultancy, Research Funding; Protagonist Therapeutics: Research Funding; Roche: Research Funding; Gilead: Research Funding; Celgene: Consultancy, Research Funding. De Stefano:Janssen Cilag: Other: Non-financial support; Celgene: Other: Non-financial support, personal fee; Bayer: Other: Non-financial support; Amgen: Other: Personal fee; Novartis: Other: Personal fee, Research Funding. Heidel:Novartis: Consultancy, Honoraria, Research Funding. Zuurman:Novartis Pharma A.G.: Current Employment. Zaiac:Novartis: Current Employment, Current equity holder in private company. Bryan:Novartis: Current Employment. Buckley:Novartis: Current Employment. Mathur:Novartis: Current Employment. Morelli:Novartis: Current Employment. Bigan:Novartis: Consultancy. Ruhl:Boston Consulting Group: Current Employment. Meier:Novartis: Consultancy.

International audience; A model of labour supply and taxation is developed in which observed hours reflect both the distribution of preferences and restrictions on choices. Hours restrictions are placed in a constrained rational choice setting in which the set of alternative hours on offer is restricted. Choices are made on a random subset of possible hours. We focus on the case where the choice set contains at most two offers. We show that when the choice set distribution is known, preferences can be identified. Conversely we show that, where preferences are known, the choice set distribution can be fully recovered. Conditions for identification of both preferences and the distribution of choice sets are also developed. We illustrate this approach in a labour supply setting with nonlinear budget constraints. Heterogeneity in the budget constraint reveals properties of the choice set. This framework is used to study the labour supply behaviour of a large sample of working age mothers in the UK, accounting for nonlinearities in the tax and welfare benefit system, fixed costs of work and restrictions on hours choices.

Introduction: Thromboembolic events (TEs) are one of the most prevalent complications in patients (pts) with polycythemia vera (PV). This real-world evidence study of the US OPTUM database evaluated the incidence of TEs in hydroxyurea (HU)-treated PV pts who either switched to ruxolitinib (RUX) after initial treatment (Tx) with HU (HU-RUX group) or continued HU Tx without switching (HU-alone group). Machine learning was then used to build a precise and scientifically robust model to predict the occurrence of TEs in PV pts with/without a history of TEs and HU failure (defined by either European LeukemiaNet [ELN] hematologic criteria or TEs). Methods: The OPTUM database comprises claims data and electronic medical records from 90 million pts (2007-2017, median stay in the database=7 years), including 69,464 PV pts. To avoid any selection bias during comparison, only pts treated prior to the RUX market launch were included in the HU-alone group (HU-RUX, n=81; HU-alone, n=195). Due to unavailability of Tx duration, time difference between the first and the last prescription was used as a proxy, and overall Tx duration was matched in both groups. TEs were assessed before Tx initiation in both groups. For HU-RUX pts, it was also assessed while on HU (median duration 27 months) and on RUX (median duration 14 months). For HU-alone pts, it was assessed during the first 27 months of Tx (any pt included in the analysis was treated for longer than this due to duration matching) and during remaining period of Tx (median duration 14 months). TEs were identified by either a restrictive definition (a list of ICD codes containing keywords from the RESPONSE study was automatically generated and manually curated) or a less restrictive one (list of ICD codes was manually expanded to include any TEs matching those from the GEMFIN study). PV pts who were exclusively treated with HU for ≥6 months were selected (n=2057) for modeling. Outcomes to be predicted were TEs in the 12 months following the end of the 6-month HU Tx period, and HU failure within 3 months of Tx. A logistic regression model was used for prediction. The baseline features extracted from the database included median lab parameters (3-6 months after HU initiation), history of thrombosis prior to primary diagnosis of PV, sociological features (age, gender), comorbidities, and concomitant medications (from inpatient/outpatient tables). Performance assessment methods included Receiver Operating Characteristic-Area Under the Curve (ROC-AUC) in early stages and confusion matrix in later stages; the findings were converted to clinically interpretable decision-tree classification algorithms. Results: Based on the extensive definition, the annual incidence of TEs in the HU-RUX and HU-alone groups, respectively, was 9% and 7% before HU initiation, which increased to 17% and 13% on HU Tx. The small difference in baseline incidence may reflect residual differences between the two groups. After a median duration of 14 months, the incidence of TEs decreased to 15% in pts who switched to RUX vs an increase to 20% in pts who continued HU Tx. A similar trend was observed using less restrictive definition (Figure 1). This definition resulted in a substantially increased incidence of TEs and a decreased predictive power of the machine-learning model. Using modeling, decision trees were developed to predict the occurrence of TEs in PV pts with/without a history of TEs. Lymphocyte percentage (13%) and platelet count (>393x103/µL) were predictors in pts with a history of TEs (Figure 2). Based on the decision tree developed to predict HU failure, phlebotomy-dependent pts with >15% RDW had a higher risk of HU failure within 3 months of Tx (Figure 3). Conclusions: A reduction in the incidence of TEs was observed in pts switching to RUX vs those who continued HU Tx. Based on the findings from this machine-learning model in PV pts, phlebotomy dependency and RDW were indicated as predictors of HU Tx failure within 3 months, whereas lymphocyte percentage+platelet count and lymphocyte percentage+RDW were predictors of incidence of TEs in pts with and without a history of TEs, respectively. Non-adjustment of the results for antiplatelet/anticoagulant Tx was a study limitation. Further validation of this machine-learning model is planned in other European databases. Disclosures Verstovsek: Celgene: Consultancy, Research Funding; Gilead: Research Funding; Promedior: Research Funding; CTI BioPharma Corp: Research Funding; Genetech: Research Funding; Protaganist Therapeutics: Research Funding; Constellation: Consultancy; Pragmatist: Consultancy; Incyte: Research Funding; Roche: Research Funding; NS Pharma: Research Funding; Blueprint Medicines Corp: Research Funding; Novartis: Consultancy, Research Funding; Sierra Oncology: Research Funding; Pharma Essentia: Research Funding; Astrazeneca: Research Funding; Ital Pharma: Research Funding. De Stefano:Celgene: Consultancy, Honoraria, Speakers Bureau; Janssen: Consultancy, Honoraria, Speakers Bureau; Amgen: Consultancy, Honoraria, Speakers Bureau; Novartis: Consultancy, Honoraria, Research Funding, Speakers Bureau; Alexion: Consultancy, Honoraria, Speakers Bureau. Heidel:Novartis: Consultancy, Honoraria, Research Funding; Celgene: Consultancy; CTI: Consultancy. Zuurman:Novartis Pharma B.V.: Employment. Zaiac:Novartis: Employment, Equity Ownership. Bigan:Novartis: Consultancy. Ruhl:Novartis: Consultancy. Meier:Novartis: Consultancy. Kiladjian:Celgene: Consultancy; Novartis: Honoraria, Research Funding; AOP Orphan: Honoraria, Research Funding.

This paper examines the determinants of the choice of the college major when the length of studies and future earnings are uncertain. We estimate a three-stage schooling decision model, focusing on the effect of expected earnings on major choice. We control for dynamic selection through the use of mixture distributions. Exploiting variations across the French business cycle in the relative returns to the majors, our results yield a very low, though significant, elasticity of major choice to expected earnings. This suggests that at least for the French university context, nonpecuniary factors are a key determinant of schooling choices. © 2011 The President and Fellows of Harvard College and the Massachusetts Institute of Technology.

This working paper has been updated to a new version, W15/02, which can be downloaded here.

A model of labour supply and taxation is developed in which observed hours reflect both the distribution of preferences and restrictions on choices. Hours restrictions are placed in a constrained rational choice setting in which the set of alternative hours on offer is restricted. Choices are made on a random subset of possible hours. We focus on the case where the choice set contains at most two offers. We show that when the choice set distribution is known, preferences can be identified. Conversely we show that, where preferences are known, the choice set distribution can be fully recovered. Conditions for identification of both preferences and the distribution of choice sets are also developed. We illustrate this approach in a labour supply setting with nonlinear budget constraints. Heterogeneity in the budget constraint reveals properties of the choice set. This framework is used to study the labour supply behaviour of a large sample of working age mothers in the UK, accounting for nonlinearities in the tax and welfare benefit system, fixed costs of work and restrictions on hours choices.

How many workers are confined indefinitely in a labor cycle, moving back and forth from temporary jobs to nonemployment periods without the potential to secure a steady job? To answer this question, we develop a mixture of four Markov chains, each of them modeling a specific type of labor market dynamics: stayers in steady jobs, stayers out of the labor market, unconfined movers who transition between states and eventually enter a permanent job, and confined movers who do not. This model is estimated using data from the French Labor Force Surveys from 2005 to 2009 for eight groups defined by gender and age to account for age-related life cycle periods. Education is included as an explanatory variable for the dynamics type. Confined movers represent 5\% of the 20- to 59-year-old individuals who have finished their studies but 12\% of 20- to 29-year-olds, and more than 25\% of 20- to 29-year-olds with less education -- up to junior high school completion. Confined movers experience much more frequent periods of unemployment than movers who can access stable jobs. Furthermore, confined movers are more likely to have less favored social backgrounds, have immigrant parents, and live in deprived areas (``Zones Urbaines Sensibles''). These empirical findings may be related to the economic mechanisms responsible for the duality of the French labor market.

This paper evaluates the French RAR program (Reseaux Ambition Reussite or Ambition Success Networks), a junior high school program started in 2006 which intended to concentrate means and funds on well-chosen disadvantaged junior high schools. We use the criteria of eligibility to estimate a regression discontinuity model. For evaluated junior high schools, the increase in per-pupil teaching hours, and the decrease in class size are disappointing. Our results also suggest that the program may have had a negative effect on teacher and pupil enrollment. Both the proportion of older teachers and the proportion of poorly qualified teachers have increased and the pupils' achievement has decreased in Grade 9, the final year of junior high school. The RAR program has increased the disparity of teachers' characteristics and of pupils' ability between schools.

A specific attention is devoted to the 55+ age group when building global labour force projections regularly updated by Insee. This requires taking into account individual heterogeneity, because the impact of pension reforms on behavior is potentially very different across individuals. For instance, increasing to 62 the minimum age of eligibility, as decided in 2010, will be neutral for people who already planned to retire after this age. It will be constraining for other people, but the impact on global labour force will depend upon employment status before retirement: postponing pension claiming for individuals who have already left the labour market does not affect this labour force, at least as long as it does not change labour market behavior before retirement age. Simulating individual transitions to retirement raises however considerable problems, to which models only bring imperfect and uncertain answers. This argues in favor of scenarized projections based on alternative assumptions. We present here results based on three options offered by the Destinie 2 microsimulation model. Depending upon these assumptions, cumulated impacts of past reforms on labour force participation rates for the 60-64 age group range, in the long run, between 10 and 40%, but starting from no-reform trends that are themselves very different from one scenario to the next. On the whole, after reforms, the long run labour force participation rate for the 60-64 group would lie within the 40%-50% bracket. This corresponds to mean ages at pension claiming lying between 64 and 65 years, and to mean ages at exit from the labour force fluctuating between 61 and 63 years, depending upon possible retroactions on career paths before full retirement.

While a new retirement pension reform is currently discussed in France, it is crucial to evaluate previous reforms. Up to now, no evaluation of the 2003 reform is available, particularly for civil servants. This article deals with the impact of this reform on the retirement behaviour of public secondary-school teachers. On the one hand, the reform has had an impact on the retirement behaviour of secondaryschool teachers who still work at 60. The probability to retire between 60 and 61 years old for those who have paid their social contributions for 37.5 years at 60 years old drops by 9 points. On the other hand, the reform seems to have changed teachers’ willingness to get the socalled “full-pension” rate. When the number of missing quarters of social contributions required to benefit from the full pension rate at 60 years old is low, the reform is not found to induce teachers born after 1944 to postpone their retirement after 61 years old. But a large number of missing quarters has still the same effect before and after the reform.

In this paper, we provide new evidence on the effect of part-time work on postsecondary educational attainment. To do so, we use samples extracted from the French Labor Force Surveys conducted over the years 1992-2002. These samples are restricted to students in initial education following university studies and preparing an Associate, a Bachelor or a Master degree. We estimate probit models with two simultaneous equations accounting for part-time working while studying and for success on the final exam, along with the decision to continue the following year in one of the models. We take the working time into account by drawing in one of the models a distinction between jobs in which more or less than 16 hours are worked per week. We use variations across departements in low-skilled youth unemployment rates and in their interactions with the father's socio-economic status in order to identify the effect of part-time work on educational attainment. Our results suggest a statistically significant and very large detrimental effect of holding a regular part-time job on graduation probability. Still, a complementary analysis shows that working while studying does not have any significant effect on the probability of continuing studies.

This paper presents some empirical evidence on the French labor market focusing on transitionsbetween stable jobs, temporary work, unemployment and non-participation. The model used is basedon a Markov chain mixture which allows one to distinguish labor market histories that are confined tocontingent work and non-employment from the non-confined ones. This enables us to identify,quantify and characterize (conditional on observable characteristics) the workers who never accede tostable jobs and remain stuck to temporary jobs and non-employment spells. We consider quarterlylabor market transitions, observed from 2003 to 2006 in the Labor Force survey (LFS). We find thaton the whole, about 5% of the working age population experience confined transition dynamics : theycannot access to stable jobs. Confined workers are less educated and are more likely to live indistressed areas.