Assuming squared error loss, we show that finding unbiased estimators and Bayes estimators can be... more Assuming squared error loss, we show that finding unbiased estimators and Bayes estimators can be treated as using a pair of linear operators that operate between two Hilbert spaces. We note that these integral operators are adjoint and then investigate some consequences of this fact.
Background. Dimension reduction methods do not always reduce their underlying indicators to a sin... more Background. Dimension reduction methods do not always reduce their underlying indicators to a single dimension. Furthermore, such methods are usually based on optimality criteria that require discarding some information. The joint probability density functions (joint pdf or JPD) can be considered as unidimensional indices. We compare such JPD indices with some traditional scoring methods.Methods. We introduced unidirectionality and co-directionality as basic requirements for a joint pdf to become an index. we then provided an argument to demonstrate its maximal informativeness property. Using all possible joint pdf conditional specifications, we estimated the JPD index. We then applied the method to two data sets: first, on the 7 Brief Pain Inventory Interference scale (BPI-I) items obtained from 8,889 US Veterans with chronic pain and, second, on a novel measure based on administrative data, the Manifestations of Psychiatric Severity Index (MoPSI), for 912 US Veterans who had appli...
Observational studies assessing causal or non-causal relationships between an explanatory measure... more Observational studies assessing causal or non-causal relationships between an explanatory measure and an outcome can be complicated by hosts of confounding measures. Large numbers of confounders can lead to several biases in conventional regression based estimation. Inference is more easily conducted if we reduce the number of confounders to a more manageable number. We discuss use of sufficient dimension reduction (SDR) summaries in estimating covariate balanced comparisons among multiple populations. SDR theory is related to the dimension reduction considered in regression theory. SDR summaries share much with sufficient statistics and encompass propensities. A specific type of SDR summary can wholly replace the original covariates with no loss of information or efficiency. Estimators with minimal expected loss can be based on these SDR summaries rather than all of the covariates.
Background: Informative response/non-response bias in survey research may skew effect size estima... more Background: Informative response/non-response bias in survey research may skew effect size estimates; however, an upper limit of tolerable bias has not been established.Objective: To use simple back-of-the-envelope calculations to estimate how much informative response/non-response bias can be introduced into a survey effort before parameter coverage is lost. Methods: We assigned attributes of military combat or sexual trauma exposure and of having/not having disability benefits for posttraumatic stress disorder (“service connection”) to a simulated population of 20,000 “Veterans.” We calculated true associations between combat and service connection and between military sexual trauma and service connection using 2X2 contingency tables and odds ratios. We simulated response/non-response bias by systematically increasing or decreasing individual cells of the 2X2 contingency tables by a factor of t between 1.01 and 2.00 (equivalent to introducing bias between 1% to 100%) across several sample sizes. We calculated new odds ratios and 95% confidence intervals for each biased contingency table, as well as the maximum possible bias for 13 response/non-response mechanisms.Results: Depending on the cell being manipulated, maximum possible response/nonresponse bias for the association between combat and service connection ranged from 10 to 260 percentage points and for the association between military sexual assault and service connection, 4 to 2,334 percentage points. For larger sample sizes (e.g., 1,000 -3,000), biases as small as 3 to 7 percentage points resulted in 95% confidence intervals that excluded the true odds ratio for some response/non-response mechanisms. Other mechanisms permitted bias of as much as 60 percentage points before parameter coverage was lost. Conclusions: Depending on which cell(s) in a 2X2 contingency table is affected, even small amounts of response/non-response bias can skew results to the point that 95% confidence intervals exclude the true odds ratio. Back-of-the-envelope techniques could help investigators identify and plan for high-risk scenarios.
We examined the interrelationships between and contributions of background, cognitive, and enviro... more We examined the interrelationships between and contributions of background, cognitive, and environmental factors to colorectal cancer (CRC) screening adherence. In this study, 2,416 average risk patients aged 50-75 from 24 Veterans Affairs medical facilities responded to a mailed survey with phone follow-up (response rate 81%). Survey data (attitudes, behaviors, demographics) were linked to facility (organizational complexity) and medical records data (diagnoses, screening history). Patients with a fecal occult blood test within 15 months, sigmoidoscopy or barium enema within 5.5 years, or colonoscopy within 11 years of the survey were considered adherent. Logistic regressions estimated the association between adherence and background, cognitive, and environmental factors. Deviance ratios examined interrelationships between factors. Population attributable risks (PAR) were used to identify intervention targets. The association of background factors with adherence was partially explained by cognitive and environmental factors. The association of environmental factors with adherence was partially explained by cognitive factors. Cognitive and environmental factors contributed equally to adherence. Factors with the highest PARs for non-adherence were age 50-64, less than two comorbidities, and lack of physician recommendation. Efforts to increase physician screening recommendations for younger, healthy patients at facilities with the lowest screening rates may improve CRC adherence in this setting.
Posttraumatic stress disorder (PTSD) is the most prevalent psychiatric condition for which vetera... more Posttraumatic stress disorder (PTSD) is the most prevalent psychiatric condition for which veterans receive service-connected disability benefits from the U.S. Department of Veterans Affairs (VA). Historically, women have been less likely than men to obtain PTSD disability benefits. The authors examined whether these gender disparities have been redressed over time and, if not, whether appropriate clinical factors account for persisting differences. This longitudinal, observational study was based on a gender-stratified, nationally representative sample of 2,998 U.S. veterans who applied for VA disability benefits for PTSD between 1994 and 1998. The primary outcome was change in PTSD service connection over a ten-year period. Forty-two percent (95% confidence interval [CI]=38%-45%) of the women and 50% (CI=45%-55%) of the men originally denied service connection for PTSD eventually received such benefits. Only 8% (CI=7%-10%) of women and 5% (CI=4%-6%) of men lost PTSD disability status. Compared with men, women had lower unadjusted odds of gaining PTSD service connection (odds ratio [OR]=.70, CI=.55-.90) and greater unadjusted odds of losing PTSD service connection (OR=1.76, CI=1.21-2.57). Adjusting for clinical factors accounted for the gender difference in gaining PTSD service connection; adjusting for clinical factors and demographic characteristics eliminated the gender difference in loss of PTSD service connection. Gender-based differences in receipt of PTSD service connection persisted in this cohort over a ten-year period but were explained by appropriate sources of variation. Further research on possible disparities in loss of PTSD disability benefits is warranted.
The location of acute ischemic infarct can affect the clinical outcome of stroke patients. We aim... more The location of acute ischemic infarct can affect the clinical outcome of stroke patients. We aimed to develop a prognostic tool based on the topographic distribution of early ischemic changes on admission computed tomography (CT) scans. Using the albumin in acute stroke (ALIAS) trials dataset, patients with anterior circulation stroke were included for analysis. A 3-month modified Rankin scale (mRs) score > 2 defined disability/death; and ≤2 defined favorable outcome. A penalized logistic regression determined independent predictors of disability/death among components of admission CT scan Alberta Stroke Program Early CT score (ASPECTS). Follow-up 24-hour CT/MRI scans were reviewed for intracranial hemorrhage (ICH). A simplified ASPECTS (sASPECTS) was developed including the caudate, lentiform nucleus, insula, and M5 components of ASPECTS-which were independent predictors of disability/death on multivariate analysis. There was no significant difference between ASPECTS and sASPECTS in prediction of disability/death (P = .738). Among patients with sASPECTS ≥ 1, the rate of favorable outcome was higher in those with intravenous (IV) thrombolytic therapy (501/837, 59.9%) versus those without treatment (91/183, 49.7%, P = .013); whereas among patients with sASPECTS of 0, IV thrombolysis was not associated with improved outcome. Also, patients with sASPECTS of 0 were more likely to develop symptomatic ICH (odds ratio = 2.62, 95% confidence interval: 1.49-4.62), compared to those with sASPECTS ≥ 1 (P = .004). Topographic assessment of acute ischemic changes using the sASPECTS (including caudate, lentiform nucleus, insula, and M5) can predict disability/death in anterior circulation stroke as accurately as the ASPECTS; and may help predict response to treatment and risk of developing symptomatic ICH.
Assuming squared error loss, we show that finding unbiased estimators and Bayes estimators can be... more Assuming squared error loss, we show that finding unbiased estimators and Bayes estimators can be treated as using a pair of linear operators that operate between two Hilbert spaces. We note that these integral operators are adjoint and then investigate some consequences of this fact.
Background. Dimension reduction methods do not always reduce their underlying indicators to a sin... more Background. Dimension reduction methods do not always reduce their underlying indicators to a single dimension. Furthermore, such methods are usually based on optimality criteria that require discarding some information. The joint probability density functions (joint pdf or JPD) can be considered as unidimensional indices. We compare such JPD indices with some traditional scoring methods.Methods. We introduced unidirectionality and co-directionality as basic requirements for a joint pdf to become an index. we then provided an argument to demonstrate its maximal informativeness property. Using all possible joint pdf conditional specifications, we estimated the JPD index. We then applied the method to two data sets: first, on the 7 Brief Pain Inventory Interference scale (BPI-I) items obtained from 8,889 US Veterans with chronic pain and, second, on a novel measure based on administrative data, the Manifestations of Psychiatric Severity Index (MoPSI), for 912 US Veterans who had appli...
Observational studies assessing causal or non-causal relationships between an explanatory measure... more Observational studies assessing causal or non-causal relationships between an explanatory measure and an outcome can be complicated by hosts of confounding measures. Large numbers of confounders can lead to several biases in conventional regression based estimation. Inference is more easily conducted if we reduce the number of confounders to a more manageable number. We discuss use of sufficient dimension reduction (SDR) summaries in estimating covariate balanced comparisons among multiple populations. SDR theory is related to the dimension reduction considered in regression theory. SDR summaries share much with sufficient statistics and encompass propensities. A specific type of SDR summary can wholly replace the original covariates with no loss of information or efficiency. Estimators with minimal expected loss can be based on these SDR summaries rather than all of the covariates.
Background: Informative response/non-response bias in survey research may skew effect size estima... more Background: Informative response/non-response bias in survey research may skew effect size estimates; however, an upper limit of tolerable bias has not been established.Objective: To use simple back-of-the-envelope calculations to estimate how much informative response/non-response bias can be introduced into a survey effort before parameter coverage is lost. Methods: We assigned attributes of military combat or sexual trauma exposure and of having/not having disability benefits for posttraumatic stress disorder (“service connection”) to a simulated population of 20,000 “Veterans.” We calculated true associations between combat and service connection and between military sexual trauma and service connection using 2X2 contingency tables and odds ratios. We simulated response/non-response bias by systematically increasing or decreasing individual cells of the 2X2 contingency tables by a factor of t between 1.01 and 2.00 (equivalent to introducing bias between 1% to 100%) across several sample sizes. We calculated new odds ratios and 95% confidence intervals for each biased contingency table, as well as the maximum possible bias for 13 response/non-response mechanisms.Results: Depending on the cell being manipulated, maximum possible response/nonresponse bias for the association between combat and service connection ranged from 10 to 260 percentage points and for the association between military sexual assault and service connection, 4 to 2,334 percentage points. For larger sample sizes (e.g., 1,000 -3,000), biases as small as 3 to 7 percentage points resulted in 95% confidence intervals that excluded the true odds ratio for some response/non-response mechanisms. Other mechanisms permitted bias of as much as 60 percentage points before parameter coverage was lost. Conclusions: Depending on which cell(s) in a 2X2 contingency table is affected, even small amounts of response/non-response bias can skew results to the point that 95% confidence intervals exclude the true odds ratio. Back-of-the-envelope techniques could help investigators identify and plan for high-risk scenarios.
We examined the interrelationships between and contributions of background, cognitive, and enviro... more We examined the interrelationships between and contributions of background, cognitive, and environmental factors to colorectal cancer (CRC) screening adherence. In this study, 2,416 average risk patients aged 50-75 from 24 Veterans Affairs medical facilities responded to a mailed survey with phone follow-up (response rate 81%). Survey data (attitudes, behaviors, demographics) were linked to facility (organizational complexity) and medical records data (diagnoses, screening history). Patients with a fecal occult blood test within 15 months, sigmoidoscopy or barium enema within 5.5 years, or colonoscopy within 11 years of the survey were considered adherent. Logistic regressions estimated the association between adherence and background, cognitive, and environmental factors. Deviance ratios examined interrelationships between factors. Population attributable risks (PAR) were used to identify intervention targets. The association of background factors with adherence was partially explained by cognitive and environmental factors. The association of environmental factors with adherence was partially explained by cognitive factors. Cognitive and environmental factors contributed equally to adherence. Factors with the highest PARs for non-adherence were age 50-64, less than two comorbidities, and lack of physician recommendation. Efforts to increase physician screening recommendations for younger, healthy patients at facilities with the lowest screening rates may improve CRC adherence in this setting.
Posttraumatic stress disorder (PTSD) is the most prevalent psychiatric condition for which vetera... more Posttraumatic stress disorder (PTSD) is the most prevalent psychiatric condition for which veterans receive service-connected disability benefits from the U.S. Department of Veterans Affairs (VA). Historically, women have been less likely than men to obtain PTSD disability benefits. The authors examined whether these gender disparities have been redressed over time and, if not, whether appropriate clinical factors account for persisting differences. This longitudinal, observational study was based on a gender-stratified, nationally representative sample of 2,998 U.S. veterans who applied for VA disability benefits for PTSD between 1994 and 1998. The primary outcome was change in PTSD service connection over a ten-year period. Forty-two percent (95% confidence interval [CI]=38%-45%) of the women and 50% (CI=45%-55%) of the men originally denied service connection for PTSD eventually received such benefits. Only 8% (CI=7%-10%) of women and 5% (CI=4%-6%) of men lost PTSD disability status. Compared with men, women had lower unadjusted odds of gaining PTSD service connection (odds ratio [OR]=.70, CI=.55-.90) and greater unadjusted odds of losing PTSD service connection (OR=1.76, CI=1.21-2.57). Adjusting for clinical factors accounted for the gender difference in gaining PTSD service connection; adjusting for clinical factors and demographic characteristics eliminated the gender difference in loss of PTSD service connection. Gender-based differences in receipt of PTSD service connection persisted in this cohort over a ten-year period but were explained by appropriate sources of variation. Further research on possible disparities in loss of PTSD disability benefits is warranted.
The location of acute ischemic infarct can affect the clinical outcome of stroke patients. We aim... more The location of acute ischemic infarct can affect the clinical outcome of stroke patients. We aimed to develop a prognostic tool based on the topographic distribution of early ischemic changes on admission computed tomography (CT) scans. Using the albumin in acute stroke (ALIAS) trials dataset, patients with anterior circulation stroke were included for analysis. A 3-month modified Rankin scale (mRs) score > 2 defined disability/death; and ≤2 defined favorable outcome. A penalized logistic regression determined independent predictors of disability/death among components of admission CT scan Alberta Stroke Program Early CT score (ASPECTS). Follow-up 24-hour CT/MRI scans were reviewed for intracranial hemorrhage (ICH). A simplified ASPECTS (sASPECTS) was developed including the caudate, lentiform nucleus, insula, and M5 components of ASPECTS-which were independent predictors of disability/death on multivariate analysis. There was no significant difference between ASPECTS and sASPECTS in prediction of disability/death (P = .738). Among patients with sASPECTS ≥ 1, the rate of favorable outcome was higher in those with intravenous (IV) thrombolytic therapy (501/837, 59.9%) versus those without treatment (91/183, 49.7%, P = .013); whereas among patients with sASPECTS of 0, IV thrombolysis was not associated with improved outcome. Also, patients with sASPECTS of 0 were more likely to develop symptomatic ICH (odds ratio = 2.62, 95% confidence interval: 1.49-4.62), compared to those with sASPECTS ≥ 1 (P = .004). Topographic assessment of acute ischemic changes using the sASPECTS (including caudate, lentiform nucleus, insula, and M5) can predict disability/death in anterior circulation stroke as accurately as the ASPECTS; and may help predict response to treatment and risk of developing symptomatic ICH.
Uploads
Papers by Siamak Noorbaloochi