Epidemiology 3
Presenter: Ntonghanwah Forcheh
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
On methods of extracting and reporting leading joint causes of death
There has been a long standing desire to extract and tabulate leading joint causes of mortality to either supplement or replace leading underlying causes for over a century. Reserach into methods for analyzing multiple causes of death (MCOD) have been intensified in recent years and various research have reported associations between two or more selected causes of death within MCOD databses. However, methods for extracting joint leading causes in a data set have remained elusive. In this paper, we show how, with a simple modification, association rules mining provides an adequate methodology to extract leading joint causes of death in a multiple cause of death data set. We apply the method to the MCOD in a middle income country. We tabulate the 10 leading multiple causes and joint leading causes of death in the data set. The age-sex mortality rates associated with these joint causes are determined and compared. Preliminary results show that the leading multiple causes of death in the country consist of a mixture of infectious and chronic diseases. The 10 leading multiple causes of death are respectively Other viral diseases (B25-B34); Human immunodeficiency virus [HIV] disease (B20-B24); Influenza and pneumonia (J10-J18); Tuberculosis (A15-A19); Other forms of heart disease (I30-I52); Aplastic and other anaemias (D60-D64); Renal failure (N17-N19); Other bacterial diseases (A30-A49); Hypertensive diseases (I10-I15) and Metabolic disorders (E70-E90). The prevalence of these causes differ by age and sex. The leading pairs of joint causes consist of other viral diseases (B25-B34) occurring together with Tuberculosis (A15-A19), Influenza and pneumonia (J10-J18), Aplastic and other anaemias (D60-D64), and Human immunodeficiency virus [HIV] disease (B20-B24). The second basket of joint pair of leading causes consist of heart diseases (I10-I15) jointly occurring with other forms of heart disease (I30-I52); Diabetes mellitus (E10-E14) and Cerebrovascular diseases (I60-I69). The top 10 pairs of leading joint causes is completed with HIV occurring with Tuberculosis (A15-A19) and with Influenza and pneumonia (J10-J18).
Epidemiology 3
Presenter: James Hanley
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
Estimating the mortality reduction produced by each round of cancer screening
In trials, or using data from regions that introduced screening programs, the mortality reduction is usually estimated using a single hazard ratio. Screening is assumed to have an immediate and sustained benefit, no matter how intense the screening. Screening targets cancers that prove fatal if detected too late (in the absence of screening, in a typical country of 5 million, c. 1000 women die of breast cancer each year, despite the treatment received when it came to attention.) Screening aims to detect and treat these fatal cancers earlier. If successful, the number of deaths despite the (screening-induced) earlier treatment will be smaller. In trials/practice, not all who die in year Y or age A will have had the same no. of screens: if aged 50 when screening was first offered, by A >= 70 they would have had 10 invitations: if aged 69 (last screening age) when it was introduced, they would have had 1. Moreover, reductions due to these will have been delayed, beginning 3-5 years after the first, and ending 4-10 years after the last screen. The reduction in Lexis cell(A,Y) is a sum of the benefits of all previous screens. If screening works as intended, the hazards are not proportional over (A,Y). Liu et al. (Int Stat Rev 83(3), 493-510, 2015) developed a model fot the reductions inside each of the (A,Y) cells. It describes the effect of one round of screening in terms of parameters for (1) when in the follow-up time the reduction is maximal (2) how large it is and (3) how dispersed in follow-up time the reductions are. Then, using the screening history in each (A,Y) cell as a design matrix, reductions over all previous screens are combined. The resulting hazard ratio curves ultimately (if follow-up extends far enough beyond the last screen) have bathtub shapes modulated by the screening histories. This model refined the results in trials of screening for colon and lung cancer. Regions where programs have been introduced (Njor et al, J Med Scr 2012;19 Suppl 1:33-41) provide the best information on mortality reductions due to breast cancer screening. We re-visit the data from the natural experiment in Funen (Njor et al, J Med Scr 2015 Mar;22(1):20-7) and apply the round-by-round model. We show how analyses of such screening data can be extended/refined to incorporate the numbers and timing of screening invitations in relation to where in the (A,Y) space the deaths do/do not occur.
Epidemiology 3
Presenter: Lixian Peng
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
Design of Primary and Sensitivity Analyses for Handling Non-future Dependence Missing Data in Clinical Trials With an Emphasis on the Type-I Error Rate Using Multiple Imputation and Pattern Mixture Model Approach
Missing data is a common problem in longitudinal clinical trials. Substantial missing data could introduce potential biases and undermine the scientific credibility of causal conclusions from clinical trials. To handle the missing data issue, it is always required by the regulatory agencies to pre-specify a primary analysis and sensitivity analysis in protocol or statistical analysis plan (SAP). Recent National Research Council (NRC) report questioned the reasonableness of the missing at random (MAR) setting as the primary analysis since MAR is a very special and doubtful assumption for the missing data mechanism, and the report encourages to use not missing at random (NMAR) setting as the primary analysis. One of the NMAR mechanisms is non-future dependence missing data (NFD-NMAR). It is also one of the recommended methods in the NRC report. This talk addresses the issue and proposes a process to investigate the mean-shift model with NFD-NMAR mechanism (NFD-Delta method). The goal is to provide, via the investigation process, a method of finding an appropriate shift parameter to specify the primary NMAR analysis in study protocol or SAP based on the criteria of maintenance of the type-I error rate for late phase trials by simulations. The simulation set-up should be based on either early phase data or information from interim data of the current trial. The shift parameter of the NFD-Delta method constitutes the sensitivity analysis. Several components were considered for the NFD shift parameter: metric/unit, magnitude, and the algorithm to place the shift to examine the effect of these components on the type-I error rate under the null hypothesis of no treatment effect. For the metric factor, four different metric units were considered: constant STD1, constant RSD1, STDk, RSDk; for the magnitude factor, different values of shift parameter f were considered to investigate which f value is the appropriate shift parameter to control the type-I error rate to the nominal level; for the algorithm to implement the delta shift, three different methods were proposed: sequential, non-sequential and single adjustment method. Extensive simulations were conducted to investigate the type-I error rate. Correctness and robustness of the results were examined.
Epidemiology 3
Presenter: Alessio Crippa
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
Point-wise averaging approach in dose-response meta-analysis of aggregated data
The traditional approach in meta-analysis of summarized data is to specify an overall functional relation (typical example is a linear trend) or a common transformation (polynomials or splines) of the quantitative predictor; apply it to each individual study, and then combine the study-specific regression coefficients using an inverse-variance weighted average. Our intent is to propose an alternative approach for meta-analysis of summarized dose-response data in order to allow for greater flexibility of the overall dose-response curve. We explore the idea of point-wise averaging of study-specific trends. The advantages of this strategy have been described in meta-analysis of individual patient data, but have never been investigated in the context of summarized data. Each individual study is allowed to have a different dose-response curve and predicted outcomes are then averaged for a specific set of values of the quantitative predictor. We will describe strengths and limitations of this new strategy and show how to present the findings in a graphical and tabular form. An empirical comparison of the dose-response curves derived from aggregated and individual patient data will be based on the SEER 9 Registries (http://seer.cancer.gov).
Epidemiology 3
Presenter: Carlo Lancia
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
A Marginal Structural Model with Dose-Delay Joint-Exposure for Assessing Reductions to Chemotherapy Intensity
We analyze chemotherapy data from a randomized trial to assess the causal effect on histological response (HR) of weakening the exposure to cytotoxic agents. Toxicities developed by patients through a chemotherapy cycle affect subsequent exposure by delaying the next cycle or reducing its dosage. HR is strongly related to the patient's past exposure to cytotoxic agents, so toxicity is at the same time a risk factor for HR and a predictor of following exposure. On the other hand, past exposure predicts toxicity levels, thus toxicity is a time-dependent confounder. Marginal Structural Models (MSMs) offer unbiased estimates of the causal effect of therapy modifications in presence of a time-dependent confounder such as toxicity. MSMs based on Inverse Probability of Treatment Weighted (IPTW) estimators create a pseudo-population by weighting each subject with the inverse of the probability of observing the allocation of a dose delay/reduction. We design a MSM to mimic a randomized trial where the reduction of the exposure intensity is no longer confounded by the toxicity. Since the intensity of the exposure can be weakened by allocating dose reductions and/or cycle delays, we propose an innovative bivariate exposure-model for the computation of the IPTW estimators. We apply this model on preoperative data from a large study in resectable osteosarcoma, EURAMOS-1, where the therapy is a multi-agent regimen called MAP (Methotrexate/Adryamicin/cisPlatin). Following clinical input, we address the following question: what is the effect on HR of reducing Methotrexate by one dose? HR is the fraction of tumor cells still viable after preoperative chemotherapy, and is recognized as a strong prognostic factor for osteosarcoma. Assessing the effectiveness of protocol modifications is a very challenging task, requiring to take care at the same time of both drugs dosages and administration schedule. In osteosarcoma adding one more agent to a 2-drug protocol significantly improves survival but the addition of a 4th drug does not; compressing a schedule is expected to achieve greater efficacy by minimizing tumor regrowth between treatment cycles. Hardly encountered in the literature, the innovative bivariate exposure we propose is here necessary to capture simultaneously reductions of the treatment intensity in dosage and schedule.
Epidemiology 3
Presenter: Amy Davidow
When: Thursday, July 14, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon C Carson Hall (Level 2)
Session Synopsis:
Quantifying the effect of spectrum bias in diagnostic accuracy studies
Spectrum bias refers to the effect of variability in the severity of patients disease on diagnostic accuracy defined by a given cut point of an underlying continuous measurement. Compared to estimates from a controlled clinical trial, specificity may be greater and sensitivity, lower, in clinical practice in the absence of strict patient exclusion/inclusion criteria. Medical literature is replete with examples of spectrum bias whereby the actual mix of patients affects diagnostic accuracy. Electrocardiographic criteria for left ventricular hypertrophy, dipstick tests for urinary tract infections, and fluorescent antinuclear antibody tests for lupus have all been empirically shown to vary in test specificity in real world settings. Here, we capture the variation in patient severity via a three stage disease model: disease free, D1 and D2. For a given diagnostic test, we develop a closed form formula which demonstrates that test specificity for D1 depends upon test sensitivity for D2 as well as the prevalence of D1 and D2. The formula readily shows that the specificity of the test for D1 is a decreasing function of the sensitivity of the test for D2. This is in addition to the well-known trade-off between sensitivity and specificity of a test for a classic two stage disease model. Using simulation, we explore the relationship between the specificity of the test for D1 and the prevalence of D1 and D2. We apply this model to tuberculosis, a pathology where diagnostic tests are frequently studied using three disease stages: susceptible, symptom-free latent tuberculosis infection, and active tuberculosis disease. Using our model, we review a published study of two commercially available diagnostic tests for latent tuberculosis infection in children and adults, the majority of whom were born in countries with high tuberculosis prevalence. Study subjects include those with active TB and controls who are a mix of susceptibles and those with latent tuberculosis infection. Our model provides an explanation of why specificity in children depended upon the way in which the control groups were constituted, an effect that was not observable among adult subjects. As there is no gold standard for latent tuberculosis infection, we use epidemiologic correlates of latent tuberculosis infection to support this explanation.
