umu.sePublications
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Methods for improving covariate balance in observational studies
Umeå University, Faculty of Social Sciences, Umeå School of Business and Economics (USBE), Statistics. (Stat4Reg)
2017 (English)Doctoral thesis, comprehensive summary (Other academic)Alternative title
Metoder för att förbättra jämförbarheten mellan två grupper i observationsstudier (Swedish)
Abstract [en]

This thesis contributes to the field of causal inference, where the main interest is to estimate the effect of a treatment on some outcome. At its core, causal inference is an exercise in controlling for imbalance (differences) in covariate distributions between the treated and the controls, as such imbalances otherwise can bias estimates of causal effects. Imbalance on observed covariates can be handled through matching, where treated and controls with similar covariate distributions are extracted from a data set and then used to estimate the effect of a treatment.

The first paper of this thesis describes and investigates a matching design, where a data-driven algorithm is used to discretise a covariate before matching. The paper also gives sufficient conditions for if, and how, a covariate can be discretised without introducing bias.

Balance is needed for unobserved covariates too, but is more difficult to achieve and verify. Unobserved covariates are sometimes replaced with correlated counterparts, usually referred to as proxy variables. However, just replacing an unobserved covariate with a correlated one does not guarantee an elimination of, or even reduction of, bias. In the second paper we formalise proxy variables in a causal inference framework and give sufficient conditions for when they lead to nonparametric identification of causal effects.

The third and fourth papers both concern estimating the effect an enhanced cooperation between the Swedish Social Insurance Agency and the Public Employment Service has on reducing sick leave. The third paper is a study protocol, where the matching design used to estimate this effect is described. The matching was then also carried out in the study protocol, before the outcome for the treated was available, ensuring that the matching design was not influenced by any estimated causal effects. The third paper also presents a potential proxy variable for unobserved covariates, that is used as part of the matching. The fourth paper then carries out the analysis described in the third paper, and uses an instrumental variable approach to test for unobserved confounding not captured by the supposed proxy variable.

Place, publisher, year, edition, pages
Umeå: Umeå universitet , 2017. , 29 p.
Series
Statistical studies, ISSN 1100-8989 ; 52
Keyword [en]
causal effect, coarsening, discretisation, proxy variables, register study, swedish social insurance agency, unobserved variables
National Category
Probability Theory and Statistics
Identifiers
URN: urn:nbn:se:umu:diva-139523ISBN: 978-91-7601-751-7 (print)OAI: oai:DiVA.org:umu-139523DiVA: diva2:1141671
Public defence
2017-10-10, Hörsal G, Humanisthuset, Umeå, 10:00 (English)
Opponent
Supervisors
Available from: 2017-09-19 Created: 2017-09-15 Last updated: 2017-10-02Bibliographically approved
List of papers
1. Data-driven Coarsening of Covariates for Causal Inference
Open this publication in new window or tab >>Data-driven Coarsening of Covariates for Causal Inference
(English)Manuscript (preprint) (Other academic)
National Category
Probability Theory and Statistics
Research subject
Statistics
Identifiers
urn:nbn:se:umu:diva-139519 (URN)
Available from: 2017-09-15 Created: 2017-09-15 Last updated: 2017-09-18
2. Proxy variables and nonparametric identification of causal effects
Open this publication in new window or tab >>Proxy variables and nonparametric identification of causal effects
2017 (English)In: Economics Letters, ISSN 0165-1765, E-ISSN 1873-7374, Vol. 150, 152-154 p.Article in journal (Refereed) Published
Abstract [en]

Proxy variables are often used in linear regression models with the aim of removing potential confounding bias. In this paper we formalise proxy variables within the potential outcomes framework, giving conditions under which it can be shown that causal effects are nonparametrically identified. We characterise two types of proxy variables and give concrete examples where the proxy conditions introduced may hold by design.

Keyword
Average treatment effect, Observational studies, Potential outcomes, Unobserved confounders
National Category
Probability Theory and Statistics
Research subject
Statistics
Identifiers
urn:nbn:se:umu:diva-129519 (URN)10.1016/j.econlet.2016.11.018 (DOI)000392568300038 ()
Available from: 2016-12-31 Created: 2016-12-31 Last updated: 2017-11-29Bibliographically approved
3. Study protocol for the evaluation of a vocational rehabilitation
Open this publication in new window or tab >>Study protocol for the evaluation of a vocational rehabilitation
Show others...
2017 (English)In: Observational Studies, Vol. 3, 1-27 p.Article in journal (Refereed) Published
Abstract [en]

This paper presents a study protocol for the evaluation of a vocational rehabilitation, namely a collaboration between the Swedish Social Insurance Agency and the Public Employment Service, where individuals needing support to regain work ability were called to a joint assessment meeting. This protocol describes a matching study design using a lasso algorithm, where we do not have access to outcome data on work ability for the treated. The matching design is based on a collection of health and socio-economic covariates measured at baseline. We also have access to a prognosis made by caseworkers on the expected length of the individual sick leave. This prognosis variable is, we argue, a proxy variable for potential unmeasured confounders. We present results showing balance achieved on observed covariates.

Keyword
Love Plots, Matching Design, Observational Study, Proxy Variable
National Category
Probability Theory and Statistics
Identifiers
urn:nbn:se:umu:diva-133847 (URN)
Available from: 2017-04-19 Created: 2017-04-19 Last updated: 2017-09-18Bibliographically approved
4. Evaluation of a Vocational Rehabilitation: Using a Proxy for Unobserved Confounders
Open this publication in new window or tab >>Evaluation of a Vocational Rehabilitation: Using a Proxy for Unobserved Confounders
Show others...
(English)Manuscript (preprint) (Other academic)
National Category
Probability Theory and Statistics
Identifiers
urn:nbn:se:umu:diva-139520 (URN)
Available from: 2017-09-15 Created: 2017-09-15 Last updated: 2017-09-18

Open Access in DiVA

fulltext(441 kB)54 downloads
File information
File name FULLTEXT01.pdfFile size 441 kBChecksum SHA-512
c15bc39cd43840d1f68dba1f0508939c074d6745f3555187f708455ae6dcd4a7be472779617781f77f3aef5067231743e27860a245ee4fdb7a51d274b55c36a9
Type fulltextMimetype application/pdf
spikblad(182 kB)11 downloads
File information
File name FULLTEXT02.pdfFile size 182 kBChecksum SHA-512
3e53ed83b448abe60017b99b9cb28eb817aea0d23ea1d2104aa439a0a89281bb2a12cddbc0d26623aebc8aba0884058b69595e5ec4737e55213e43d4b966b80d
Type spikbladMimetype application/pdf

Search in DiVA

By author/editor
Fowler, Philip
By organisation
Statistics
Probability Theory and Statistics

Search outside of DiVA

GoogleGoogle Scholar
Total: 65 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

isbn
urn-nbn

Altmetric score

isbn
urn-nbn
Total: 555 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf