Identification Of Treatment Effects Research Articles

Item-level heterogeneous treatment effects of selective serotonin reuptake inhibitors (SSRIs) on depression: implications for inference, generalizability, and identification

Abstract Objectives In analysis of randomized controlled trials (RCTs) with patient-reported outcome measures (PROMs), Item Response Theory (IRT) models that allow for heterogeneity in the treatment effect at the item level merit consideration. These models for “item-level heterogeneous treatment effects” (IL-HTE) can provide more accurate statistical inference, allow researchers to better generalize their results, and resolve critical identification problems in the estimation of interaction effects. In this study, we extend the IL-HTE model to polytomous data and apply the model to determine how the effect of selective serotonin reuptake inhibitors (SSRIs) on depression varies across the items on a depression rating scale. Methods We first conduct a Monte Carlo simulation study to assess the performance of the polytomous IL-HTE model under a range of conditions. We then apply the IL-HTE model to item-level data from 24 RCTs measuring the effect of SSRIs on depression using the 17-item Hamilton Depression Rating Scale (HDRS-17) and estimate heterogeneity by subscale (HDRS-6). Results Our simulation results show that ignoring IL-HTE can yield standard errors that are as much as 50 % too small and create significant bias in treatment by covariate interaction effects when item-specific treatment effects are correlated with item location, and that the application of the IL-HTE model resolves these issues. Our empirical application shows that while the average effect of SSRIs on depression is beneficial (i.e., negative) and statistically significant, there is substantial IL-HTE, with estimates of the standard deviation of item-level effects nearly as large as the average effect. We show that this substantial IL-HTE is driven primarily by systematically larger effects on the HDRS-6 subscale items. Conclusions The IL-HTE model has the potential to provide new insights for the inference, generalizability, and identification of treatment effects in clinical trials using PROMs.

Treatment effect identification using two-level designs with partially ignorable missing data

Missingness is often present in treatment effect inference studies with response datasets. However, the existing methods could not deal with the partially ignorable missingness (PIM) of response data from many time points. These data often have two characteristics: (a) missing in the framework of counterfactual outcomes; (b) missing responses arising from dropouts. One challenge is to identify the treatment effect in three-way arrays of outcomes with missing data.Thus, this paper proposes a novel method, PIM estimator (PIME), for inferring the differential effect of treatment. This method aims to estimate the average treatment effect using the three-way outcome array of a two-level design. This method can identify the selection bias and quantify it with the difference in outcome means between complete and incomplete cases. With the sensitivity analysis of dropout fractions, the results suggest there exists a minimum point of the upper bound of the difference estimation with the specified models. The results with real applications show that the proposed method is efficient for the treatment effect estimation. The proposed method was compared with the related methods (with dropouts imputed or removed). The results indicate that patients with higher blood glucose levels are more likely to drop out than the others.

Identifying Effects of Multiple Treatments in the Presence of Unmeasured Confounding

Identification of treatment effects in the presence of unmeasured confounding is a persistent problem in the social, biological, and medical sciences. The problem of unmeasured confounding in settings with multiple treatments is most common in statistical genetics and bioinformatics settings, where researchers have developed many successful statistical strategies without engaging deeply with the causal aspects of the problem. Recently there have been a number of attempts to bridge the gap between these statistical approaches and causal inference, but these attempts have either been shown to be flawed or have relied on fully parametric assumptions. In this article, we propose two strategies for identifying and estimating causal effects of multiple treatments in the presence of unmeasured confounding. The auxiliary variables approach leverages variables that are not causally associated with the outcome; in the case of a univariate confounder, our method only requires one auxiliary variable, unlike existing instrumental variable methods that would require as many instruments as there are treatments. An alternative null treatments approach relies on the assumption that at least half of the confounded treatments have no causal effect on the outcome, but does not require a priori knowledge of which treatments are null. Our identification strategies do not impose parametric assumptions on the outcome model and do not rest on estimation of the confounder. This article extends and generalizes existing work on unmeasured confounding with a single treatment and models commonly used in bioinformatics. Supplementary materials for this article are available online.

Identifying treatment effects in the presence of confounded types

In this paper, I consider identification of treatment effects when the treatment is endogenous. The use of instrumental variables is a popular solution to deal with endogeneity, but this may give misleading answers when the instrument is invalid. I show that when an (unobserved) instrument is invalid due to correlation with the first stage unobserved heterogeneity, a proxy for the instrument helps partially identify not only the local average treatment effect, but also the entire potential outcomes distributions for compliers. I exploit the fact that the distribution of the observed outcome in each group defined by the treatment and the instrument is a mixture of the distributions of interest. I write the identified set in the form of conditional moment inequalities, and provide an easily implementable inference procedure. Under some tail restrictions, the potential outcomes distributions are point-identified for compliers. Finally, I illustrate my methodology on data from the National Longitudinal Survey of Young Men to estimate returns to college using college proximity as a proxy for the instrument low college cost. I find that a college degree increases the average hourly wage of the compliers by 15%–30%.

ISS stage and network risk score to predict benefits of multiple myeloma treatment options.

e20511 Background: Multiple myeloma (MM) is molecularly heterogeneous with many treatment regimens developed targeting different aspects of the disease. Multi-omics studies are now widely available, facilitating a precision medicine approach for personalized MM treatments. However, data from these studies vary with respect to treatment regimens, patient age, disease stage, and genomic alterations, thus complicating identification of treatment effects. Methods: From a Multiple Myeloma Molecular Causal Network (M3CN) model we previously described, we identified a prognostic subnetwork (prognNET) associated with survivals of newly diagnosed MM (NDMM) patients. We derived a network risk score (NRS) for each patient by inferring their molecular state with respect to prognNET, and partitioned patients into NRS low, medium, and high groups (referred as NRSL/M/H). We validated NRSL/M/H to be significantly associated with progression free survival (PFS) and overall survival (OS) in single variate and multi-variate Cox regression models including age and ISS stage in TT2,3,4 and HOVON studies. Results: To determine the impact of NRS in an independent contemporary dataset, we examined NDMM patients with RNAseq data in the MMRF-CoMMpass study. Among young patients (age < 65), ones with combined therapies (CoTs) as the first line therapy (n = 363) had better survival than ones on proteasome inhibitors (PIs) (n = 93) (p = 0.027 and 0.002 for PFS and OS). When stratified by ISS stage and NRS, there was no significant difference between CoTs and PIs except that in the stage I-NRSL group, patients on PIs had better PFS than ones on CoTs (p = 0.016). This suggests that NRS may be able to discriminate ISS stage I patients that do not need CoT. Similarly, we studied the ability of NRS to discriminate patients that benefit from autologous stem cell transplant (ASCT). In a model combining age, ISS stage, NSR, and induction treatment, ASCT was associated with better survival. When stratified young patients on CoTs by ISS stage and NRS, only the stage II-NRSM and stage III-NRSM patients benefited from ASCT in term of PFS (p = 0.01 and 0.005) and OS (p = 0.004 and 0.04). Conclusions: Stage and NRS group should be considered in personalizing treatment options for NDMM patients.

Identification of Treatment Effects with Mismeasured Imperfect Instruments

In this article, I develop a novel identification result for estimating the effect of an endogenous treatment using a proxy of an unobserved imperfect instrument. I show that the potential outcomes distributions are partially identified for the compliers. Therefore, I derive sharp bounds on the local average treatment effect. I write the identified set in the form of conditional moments inequalities, which can be implemented using existing inferential methods. I illustrate my methodology on the National Longitudinal Survey of Youth 1979 to evaluate the returns to college attendance using tuition as a proxy of the true cost of going to college. I find that the average return to college attendance for people who attend college only because the cost is low is between 29% and 78%.

Identification of treatment effects with selective participation in a randomized trial

Randomized trials (RTs) are used to learn about treatment effects. This paper studies identification of average treatment response (ATR) and average treatment effect (ATE) from RT data under various assumptions. The focus is the problem of external validity of the RT. RT data need not point identify the ATR or ATE because of selective participation in the RT. The paper reports partial‐identification and point‐identification results for the ATR and ATE based on RT data under a variety of assumptions. The results include assumptions sufficient to point identify the ATR or ATE from RT data. Under weaker assumptions, the ATR or ATE is partially identified. Further, attention is given to identification of the sign of the ATE and identification of whether participation in the RT is selective. Finally, identification from RT data is compared to identification from observational data.

Identifying Treatment Effects in the Presence of Confounded Types

In this paper, I consider identification of treatment effects when the treatment is endogenous. The use of instrumental variables is a popular solution to deal with endogeneity, but this may give misleading answers when the instrument is invalid. I show that when an (unobserved) instrument is invalid due to correlation with the first stage unobserved heterogeneity, a proxy for the instrument helps partially identify not only the local average treatment effect, but also the entire potential outcomes distributions for compliers. I exploit the fact that the distribution of the observed outcome in each group defined by the treatment and the instrument is a mixture of the distributions of interest. I write the identified set in the form of conditional moment inequalities, and provide an easily implementable inference procedure. Under some tail restrictions, the potential outcomes distributions are point-identified for compliers. Finally, I illustrate my methodology on data from the National Longitudinal Survey of Young Men to estimate returns to college using college proximity as a proxy for the instrument low college cost. I find that a college degree increases the average hourly wage of the compliers by 15-30%.

Bounds on treatment effects on transitions

This paper considers the identification of treatment effects on conditional transition probabilities. We show that even under random assignment only the instantaneous average treatment effect is point identified. Since treated and control units drop out at different rates, randomization only ensures the comparability of treatment and controls at the time of randomization, so that long-run average treatment effects are not point identified. Instead, we derive bounds on these average effects. Our bounds do not impose (semi)parametric restrictions, for example, proportional hazards. We also explore assumptions such as monotone treatment response, common shocks and positively correlated outcomes that tighten the bounds.

Treatment-effect identification without parallel paths

Abstract Imagine a region suffering from a widening income gap that becomes eligible for a generous transfer programme (the treatment). Imagine difference-in-differences analysis (DD) — a before-and-after comparison of the income-level difference — shows that the handicap has risen. Most observers would conclude to the policy’s inefficiency. But second thoughts are needed, because DD rests heavily on the validity of a key assumption: parallel paths in the absence of treatment; an assumption that is often violated. To cope with this problem, economists traditionally include polynomial (linear, quadratic…) trends among the regressors, and estimate the treatment effect as a once-in-a-time trend shift. In practice that strategy does not work very well, because inter alia the estimation of the trend uses post-treatment data. What is needed is a method that i) uses pre-treatment observations to capture linear or non-linear trend differences, and ii) extrapolates these to compute the treatment effect. This paper shows how this can be achieved using a fully-flexible version of the canonical DD equation. It also contains an illustration using data on a 1994–2006 EU programme that was implemented in the Belgian province of Hainaut.

Identification of Treatment Effects Under Conditional Partial Independence

Conditional independence of treatment assignment from potential outcomes is a commonly used but nonrefutable assumption. We derive identified sets for various treatment effect parameters under nonparametric deviations from this conditional independence assumption. These deviations are defined via a conditional treatment assignment probability, which makes it straightforward to interpret. Our results can be used to assess the robustness of empirical conclusions obtained under the baseline conditional independence assumption.

Treatment-Effect Identification Without Parallel paths. An Illustration in the Case of Objective 1 - Hainaut/Belgium, 1994-2006

Imagine an impoverished region that becomes eligible for a generous transfer programme (the treatment). Imagine difference-in-differences analysis (DD) — a before-and-after comparison of the income-level handicap — shows that the handicap has risen. Most observers would conclude to the policy's inefficiency. The point made in this paper is that second thoughts are needed, because DD rests heavily on the validity of a key assumption: parallel paths in the absence of treatment. What is more, when several pre-treatment periods are available in the data, it can easily be assessed and, if necessary, abandoned in favour of more relevant ones.

Alternative diff-in-diffs estimators with several pretreatment periods

ABSTRACTThis paper deals with the identification of treatment effects using difference-in-differences estimators when several pretreatment periods are available. We define a family of identifying nonnested assumptions that lead to alternative difference-in-differences estimators. We show that the most usual difference-in-differences estimators imply equivalence conditions for the identifying nonnested assumptions. We further propose a model that can be used to test multiple equivalence conditions without imposing any of them. We conduct a Monte Carlo analysis and apply our approach to several recent papers to show its practical relevance.

Identification of Treatment Effects from Synthetic Controls

This paper considers the Generalized Synthetic Control method of Xu (2015). It presents a characterization of the identified set for this model. In addition it presents a version in which the matrix factorization is unique without strong restrictions on the matrix factors. The approach assumes a panel data set represented by a linear factor model with no structural breaks other than the treatment for the treated unit(s). The method calculates the unit specific treatment effect using matrix factorization of the pre-treatment panel data. The paper numerically determines the identified set of the average treatment on the treated. Statistical uncertainty in the size of the treatment effect is calculated by boot-strapping the errors using the observed residuals from linear factor model estimated in the pre-treatment period. Like Xu (2015) the approach assumes that the researcher has a large number of pre-treatment periods. The estimator and identification results for the matrix factorization is presented in Adams (2016b). The approach is illustrated using synthetic data from Xu (2015), the treatment effect of electoral law changes and country level data from OECD countries before and after the reunification of Germany. The analysis presents bounds on the average treatment effect that do not include zero but do include the actual average treatment effect in the synthetic data. It estimates bounds on the average treatment on the treated for same day voter registration that includes zero and shows that reunification systematically decreased the output of the former West Germany.

Identification of Treatment Effects under Conditional Partial Independence

Conditional independence of treatment assignment from potential outcomes is a commonly used but nonrefutable assumption. We derive identified sets for various treatment effect parameters under nonparametric deviations from this conditional independence assumption. These deviations are defined via a conditional treatment assignment probability, which makes it straightforward to interpret. Our results can be used to assess the robustness of empirical conclusions obtained under the baseline conditional independence assumption.

How Often Is Treatment EffectIdentified in Axillary Nodes with a Pathologic Complete Response After Neoadjuvant Chemotherapy?

False-negative rates (FNR) of sentinel node biopsy (SNB) after neoadjuvant chemotherapy (NAC) in node-positive (N+) breast cancer patients are <10% when ≥3 negative SNs are obtained. Marking positive nodes has been suggested to reduce FNR. Identification of treatment effect in the nodes post-NAC is an alternative to decrease FNR. We evaluated the frequency of treatment effect in N+ patients after a pathologic complete response (pCR) with NAC. Biopsy-proven N+ patients receiving NAC were identified. Patients with nodal pCR after axillary lymph node dissection (ALND) or SNB with dual mapping and ≥3 SNs removed were evaluated for treatment effect; ALND and SNB patients were compared. From January 2009 to December 2015, 528 N+ patients received NAC. Of these, 204 had a nodal pCR, 135 had an ALND, and 69 had SNB. Median age was 49years, 15% were hormone receptor positive (HR+)/HER2-, 27% triple negative, and 58% HER2+. The median number of nodes removed in ALND patients was 17 versus 4 in SNB patients. Treatment effect in nodes was identified in 192 patients (94%) and was more common in ALND versus SNB patients (97 vs 88%; p=.02). HR+ patients and patients without a breast pCR were less likely to have treatment effect in the nodes (p=.05). Other characteristics did not differ. Following NAC, SNs with treatment effect were retrieved in 88% of patients without marking nodes, suggesting that nodal clipping may not be necessary to achieve an acceptable FNR. Longer follow-up is needed to determine regional recurrence rates in the SN-only cohort.

Cost savings of developmental screenings: Evidence from a nationwide program

Early intervention is considered the optimal response to developmental disorders in children. We evaluate a nationwide developmental screening program for preschoolers in Austria and the resulting interventions. Identification of treatment effects is determined by a birthday cutoff-based discontinuity in the eligibility for a financial incentive to participate in the screening. Assigned preschoolers are 14.5 percentage points more likely to participate in the program. For participants with high socio-economic status (SES), we find little evidence for interventions and consistently no effect on healthcare costs in the long run. For low SES preschoolers, we find evidence for substantial interventions, but only weak evidence for cost savings in the long run.

Identification of treatment effects under imperfect matching with an application to Chinese elite schools

This paper extends the treatment effect framework for causal inference to contexts in which the instrument appears in a data set that can only be linked imperfectly to the treatment and outcome variables contained in another data set. To overcome this problem, I form all pairwise links between information on the instrument and information on the treatment and outcome matched by the commonly recorded personal characteristics in both data sets. I show how these imperfect conditional matches can be used to identify both the average and distributional treatment effects for compliers of the common units of the two data sets. This multiple data source approach is then applied to analyze the effect of attending an elite middle school in a Chinese city where schools' admissions lottery records can only be linked imperfectly to the administrative student records.

Bounds on Treatment Effects on Transitions

This paper considers identification of treatment effects on conditional transition probabilities. We show that even under random assignment only the instantaneous average treatment effect is point identified. Because treated and control units drop out at different rates, randomization only ensures the comparability of treatment and controls at the time of randomization, so that long run average treatment effects are not point identified. Instead we derive informative bound on these average treatment effects. Our bounds do not impose (semi)parametric restrictions, as e.g. proportional hazards. We also explore various assumptions such as monotone treatment response, common shocks and positively correlated outcomes that tighten the bounds.

Does Marriage Boost Men’s Wages?: Identification of Treatment Effects in Fixed Effects Regression Models for Panel Data

Social scientists have generated a large and inconclusive literature on the effect(s) of marriage on men’s wages. Researchers have hypothesized that the wage premium enjoyed by married men may reflect both a tendency for more productive men to marry and an effect of marriage on productivity. To sort out these explanations, researchers have used fixed effects regression models for panel data to adjust for selection on unobserved time-invariant confounders, interpreting coefficients for the time-varying marriage variables as effects. However, they did not define these effects or give conditions under which the regression coefficients would warrant a causal interpretation. Consequently, they failed to appropriately adjust for important time-varying confounders and misinterpreted their results. Regression models for panel data with unobserved time-invariant confounders are also widely used in many other policy-relevant contexts and the same problems arise there. This article draws on recent statistical work on causal inference with longitudinal data to clarify these problems and help researchers use appropriate methods to model their data. A basic set of treatment effects is defined and used to define derived effects. Causal models for panel data with unobserved time-invariant confounders are defined and the treatment effects are reexpressed in terms of these models. Ignorability conditions under which the parameters of the causal models are identified from the regression models are given. Even when these hold, a number of interesting and important treatment effects are typically not identified.