Skip to content

About

Doctoral thesis research as DataOps & DevOps.

Resources

Readme
License
Activity

Stars

1 star

Watchers

1 watching

Forks

0 forks
Report repository
BranchesTags

Repository files navigation

“If You Only Knew the Power of the Dark Side”: Examining Within-Person Fluctuation in Psychological Need Frustration, Burnout, and Turnover Intentions Across a Workday.

Author & Maintainer: Demetrius K. Green

Email | github pic GitHub | linkedin pic LinkedIn | ResearchGate

Results Dashboard: https://esm-study-dashboard.netlify.app/

A dissertation submitted to the Graduate Faculty of North Carolina State University in partial fulfillment of the requirements for the degree of Doctor of Philosophy.

Industrial-Organizational Psychology

Raleigh, North Carolina (2026)

Technical Setup

This project combines academic research, software engineering, data engineering, and advanced statistical analysis. The Makefile is the primary interface — make help lists everything.

Windows is not supported. The R parallel backend and bash wrapper scripts require macOS or Linux.

Prerequisites

Install these tools before cloning. None are included in the repository.

Linux users: The table below gives apt-based hints, but R installation requires additional steps (signing key, CRAN PPA, system build libraries). gcp/deploy/setup_gcp_vm.sh contains a complete, tested R installation sequence for Ubuntu 22.04 and is the best reference for the R section — adapt it for your distro as needed. A dedicated Linux developer guide may be added as the project matures.

Tool Purpose macOS Linux
git Version control Xcode CLT (xcode-select --install) apt install git
make Task runner Xcode CLT apt install build-essential
pyenv Python version manager — reads .python-version (pins 3.12.11) brew install pyenv + shell init pyenv-installer
direnv Auto-loads .envrc on cd brew install direnv + shell hook apt install direnv + shell hook
uv Python dependency manager + virtualenv brew install uv or curl -LsSf https://astral.sh/uv/install.sh | sh curl -LsSf https://astral.sh/uv/install.sh | sh
R ≥ 4.4 Statistical analysis CRAN macOS binary CRAN PPA

Quick Start

These five steps bring both the Python and R environments to a fully working state.

1. Clone and enter the repo

git clone <repo-url>
cd dkg-phd-thesis

2. Set up environment variables

cp .env.example .env
# Edit .env if your GCP project ID differs from 'dkg-phd-thesis'

3. Allow direnv (Python environment)

direnv allow

This triggers .envrc, which:

  • Installs Python 3.12.11 via pyenv (if not already present)
  • Validates uv.lock against pyproject.toml (uv lock --check)
  • Runs uv sync --frozen into .venv/
  • Loads .env variables into the shell
  • Activates the virtualenv

The message direnv: error .envrc is blocked on first entry is normal — just run direnv allow.

3a. (iCloud users) R package cache

uvr's global cache lives at ~/.uvr/ (local disk, never on iCloud). The project library at .uvr/library/ is on iCloud if the project is in iCloud Drive, but if evicted, uvr sync restores it from the local cache in seconds. No user action is required.

4. Run full setup (R environment + git hooks)

make setup

This runs three targets in sequence:

  • setup_r — installs R packages from uvr.toml via uvr sync
  • setup_python — re-validates uv.lock and runs uv sync (idempotent after step 3)
  • setup_hooks — symlinks scripts/hooks/pre-commit into .git/hooks/

Note on pre-commit linting: the repo's pre-commit hook delegates R (lintr) and Python (ruff) lint checks to ~/.claude/hooks/repo-pre-commit.sh, which is part of the author's dotfiles. Without that dispatcher in place, staged-file linting is skipped silently (each script exits gracefully if its tool or dispatcher is absent). The lock file guard runs regardless. Collaborators without the dotfiles will not see linting blocked at commit time and should run make py_lint and lintr::lint_dir() manually before pushing.

5. Verify

make validate   # R directory structure check (fast, no internet)
make py_test    # Python test suite (fully mocked, no GCP credentials needed)
make py_lint    # Ruff linting + format check

All three should pass on a clean setup. If make validate fails, confirm Rscript --version shows R ≥ 4.4. If make py_test fails, confirm direnv allow completed without errors and which python points to .venv/bin/python.

Tracks

After Quick Start, run whichever tracks are relevant to your work.

Track A — R Analysis (Power Analysis + Synthetic Data)

Requires Quick Start steps 1–4.

make power_analysis_dev    # Power analysis: 3 combos × 10 sims (seconds)
make power_analysis_prod   # Power analysis: full grid, runs in background (hours)
make power_visual          # Power curve figures (after any run)
make synthetic_analysis    # Synthetic data: EDA → correlation → measurement → MLM

See analysis/run_power_analysis/README.md for the full parameter grid and configuration. See analysis/run_synthetic_data/README.md for synthetic data inputs and outputs.

Study data (the official pipeline on real participant data from BigQuery):

make study_export         # Rebuild BQ fact tables, re-export CSVs (prompts for confirmation)
make study_data_quality   # 1. Careless responding screening -> cleaned CSV
make study_eda            # 2. Exploratory data analysis
make study_correlation    # 3. Correlation analysis
make study_measurement    # 4. Measurement model (CFA)
make study_mlm            # 5. Multilevel model (main analysis)
make study_analysis       # Steps 1-5 in sequence (study_data_quality through study_mlm)
make study_tables         # 6. Publication-ready Word tables
make study_all            # Full study pipeline: export -> steps 1-5 -> tables (guaranteed fresh)

See analysis/run_study_analysis/README.md for the data pipeline, careless-responding screening, and variable definitions.

Track B — Python Development (Testing, Linting, Local Dev Server)

Requires Quick Start steps 1–3. No GCP credentials needed — all tests are fully mocked.

make py_test       # Run test suite
make py_lint       # Ruff check + format check
make py_format     # Auto-fix formatting
make py_sqlfmt     # SQL formatting check
make gcp_dev       # Local dev server on :8080 (default: run-qualtrics-scheduling)
make gcp_dev FN=run-intake-confirmation   # Override function

See gcp/README.md for architecture details, schema change workflow, and cURL testing.

Track C — GCP Deployment

Requires Track B + gcloud CLI authenticated to dkg-phd-thesis + Twilio secrets in Secret Manager. See gcp/README.md → "Secrets" section for credential setup.

Deploy in this order (resources depend on each other):

make gcp_infra_up                          # BigQuery tables
make gcp_pubsub_up                         # Pub/Sub topics
make gcp_deploy FN=run-qualtrics-scheduling
make gcp_deploy FN=run-intake-confirmation
make gcp_deploy FN=run-followup-scheduling
make gcp_deploy FN=run-followup-response
make gcp_gateway_up                        # API Gateway (last — needs functions deployed)

End-to-end test:

make gcp_gateway_test         # Fixed survey times
make gcp_gateway_test NOW=1   # Schedule at now + 16/32/48 min

See make help_gcp for the full GCP command reference.

Track D — GCP VM for Large Power Analysis

Requires Track B + gcloud CLI. The VM is Linux-only and is used solely for R simulations.

make gcp_compute_up      # Create VM (c3-highcpu-176)
make gcp_compute_status  # Show VM state + external IP
make gcp_compute_ssh     # SSH in

On the VM (run once after first SSH):

cd dkg-phd-thesis
bash gcp/deploy/setup_gcp_vm.sh            # Install R, system libs, uvr packages

Run the analysis:

make power_analysis_gcp_benchmark          # Timing probe (run this first)
nohup make power_analysis_gcp_prod &       # Full grid in background (3,645 cells)

After completion:

make gcp_compute_scp     # Download results to local machine
make gcp_compute_down    # Delete VM to stop billing

See analysis/run_power_analysis/README.md for benchmarks and troubleshooting.

Track E: Results Dashboard

Results from the study analysis are presented as an interactive dashboard, live at https://esm-study-dashboard.netlify.app/. It is a static, client-side app deployed on Netlify from dashboard/; there is no setup track to run here beyond what Quick Start already provides. See dashboard/README.md for what it shows, how its data is built, and how to run or deploy it.

Dependency Management

Both uv.lock and uvr.lock are frozen against accidental changes:

  • A pre-commit hook blocks commits that stage either lock file
  • Bypass for intentional updates: ALLOW_LOCK_COMMIT=1 git commit ...
  • Python update path: uv lock → uv sync → commit with bypass
  • R update path: uvr lock → commit with bypass

Troubleshooting

Symptom Fix
direnv: error .envrc is blocked Run direnv allow
uv.lock is out of sync Run uv lock
uvr sync fails Run uvr doctor to diagnose; check uvr.toml for invalid specs
Tests fail with import errors Confirm direnv allow completed; check which python points to .venv/bin/python
Rscript not found or R version < 4.4 Install R ≥ 4.4 from CRAN; verify with Rscript --version
make validate fails with path error Confirm you are running from the project root

For a full command reference: make help (all commands) or make help_gcp (GCP-specific).

Table of Contents

Overview

This study utilizes an experience sampling design with three within-day measurement occasions across a single workday to model need frustration depletion and burnout trajectories as they relate to momentary turnover intentions. Grounded in Self-Determination Theory (SDT) and Basic Psychological Needs Theory (BPNT), the study examines how the active frustration of autonomy, competence, and relatedness needs — as well as the accumulated energetic depletion reflected in burnout — predicts shifts in turnover intentions within a workday. A multilevel (mixed-effects) modeling paradigm differentially parses and evaluates within-person (Level 1) and between-person (Level 2) effects, with psychological contract breach/violation and job satisfaction entered as stable Level 2 predictors, and meeting load examined as a within-person moderator.

Conceptual Framework

Participants will be measured across three timepoints within a single respective workday. There will be specific item-stems that corresponds to a particular timepoint: (1) Since starting work today, (2) Since the first survey check-in, and (3) Since the last survey check-in. Items within each variable-block will be randomized to reduce order bias.

Level 1 Variables (L1): time-varying | Level 2 Variables (L2): time-invariant

Criterion Variable:

  • Turnover Intentions (L1)

Predictor Variables:

  • Burnout (L1)
  • Need Frustration (L1)
  • Meeting Load (L1)
  • Psychological Contract (L2)
  • Job Satisfaction (L2)

Model Covariates (all collected; age, tenure, and positive/negative affect included as model controls):

  • Positive & Negative Affect (L2)
  • Age (L2)
  • Job Tenure (L2)
  • Ethnicity (L2, collected, not modeled)
  • Gender (L2, collected, not modeled)
  • Education Level (L2, collected, not modeled)
  • Remote Flag (L2, collected, not modeled)

Marker Variable:

  • Attitude Toward the Color Blue (L1)

Hypotheses

  • H1a. Holding all other within-person predictors constant, individuals who experience higher levels of need frustration relative to their personal average at a given measurement occasion are expected to report stronger turnover intentions.

  • H1b. Individuals who, on average, experience higher levels of need frustration are expected to report stronger turnover intentions, controlling for all other between-person means and within-person deviations.

  • H2a. Holding all other within-person predictors constant, individuals who experience higher levels of burnout relative to their personal average at a given measurement occasion are expected to report stronger turnover intentions.

  • H2b. Individuals who, on average, experience higher levels of burnout are expected to report stronger turnover intentions, controlling for all other between-person means and within-person deviations.

  • H3a. Holding all other within-person predictors and between-person means constant, individuals who have more meetings than their personal average at a given measurement occasion are expected to report higher levels of turnover intentions; meeting count will positively moderate the relationships between need frustration and turnover intentions and between burnout and turnover intentions.

  • H3b. Holding all other within-person predictors and between-person means constant, individuals who spend more time in meetings than their personal average at a given measurement occasion are expected to report higher levels of turnover intentions; meeting minutes will positively moderate the relationships between need frustration and turnover intentions and between burnout and turnover intentions.

  • H4a. On average, turnover intentions are expected to be positively associated with individuals who experience chronically higher levels of psychological contract breach.

  • H4b. On average, turnover intentions are expected to be positively associated with individuals who experience chronically higher levels of psychological contract violation.

  • H5. On average, turnover intentions are expected to be negatively associated with individuals who report consistently higher levels of job satisfaction.

Research Model

thesis-research-model

Measures

Level 1 Variables (L1)

Shirom-Melamed Burnout Measure (SMBM; Shirom & Melamed, 2006)

{Timepoint-stem}, indicate how often you have experienced the following feelings or behaviors?

Response Anchors: Frequency-based (see Tong et al., 2020 for vague vs. precise justification)

  • 1 = Never
  • 2 = Once
  • 3 = Twice
  • 4 = Three times
  • 5 = More than three times
Expand for items

Physical Fatigue

  1. I have felt tired.
  2. I have been lacking energy for my work.
  3. I have felt physically drained.
  4. I have felt fed-up.
  5. I have felt like my “batteries” are “dead.”
  6. I have felt burned out.

Cognitive Weariness

  1. My thinking process has been slow.
  2. I have had difficulty concentrating.
  3. I have felt like I am not thinking clearly.
  4. I have felt like I am not focused on my thinking.
  5. I have had difficulty thinking about complex things.

Emotional Exhaustion

  1. I feel I have been unable to respond sensitively to the needs of coworkers or customers.
  2. I feel I have been incapable of investing emotionally in coworkers or customers.
  3. I feel I have been incapable of being sympathetic to coworkers or customers.

Basic Psychological Needs Thwarting Scale (PNTS; Bartholomew et al., 2011)

{Timepoint-stem}, indicate the extent to which you agree or disagree with the following statements about your experiences so far at work.

Response Anchors: Agreement-based

  • 1 = Strongly disagree
  • 2 = Somewhat disagree
  • 3 = Neither agree nor disagree
  • 4 = Somewhat agree
  • 5 = Strongly agree
Expand for items

Competency Thwarting

  1. I have felt incompetent because others have put unrealistic expectations on me.
  2. I have been told things that made me feel incompetent.
  3. There have been situations where I was made to feel inadequate.
  4. I have felt inadequate because I have not been given opportunities to fulfil my potential.

Autonomy Thwarting

  1. I have felt prevented from making choices regarding the way I approach my work.
  2. I have felt pushed to behave in certain ways.
  3. I have felt obliged to follow work decisions made for me.
  4. I have felt under pressure to agree with the work regimen provided to me.

Relatedness Thwarting

  1. I have felt rejected by those around me.
  2. I have felt like others are dismissive of me.
  3. I have felt like other people dislike me.
  4. I have felt like some of my coworkers are envious of my work or ideas.

Attitude Toward the Color Blue (ATCB; Miller et al., 2024; Miller & Simmering, 2023)

{Timepoint-stem}, indicate the extent to which you agree or disagree with the following statements.

Response Anchors: Agreement-based

  • 1 = Strongly disagree
  • 2 = Somewhat disagree
  • 3 = Neither agree nor disagree
  • 4 = Somewhat agree
  • 5 = Strongly agree
Expand for items
  1. Blue is a lovely color
  2. Blue is a nice color
  3. I think blue is a pretty color
  4. I like the color blue

Meeting Load

Expand for items

  1. {Timepoint-stem}, how many meetings have you attended?

    • Response Anchors: Count-based (i.e., integer-type)
      • 0 to n

  2. {Timepoint-stem}, how much time (in minutes) have you spent in meetings?

    • Response Anchors: Continuous
      • 0 to n

Turnover Intentions (Spector et al., 1988)

Note: See Matthews et al., 2022 for theoretical justification of single item measures.

{Timepoint-stem}, how often have you thought about leaving your current employer?

Response Anchors: Frequency-based

  • 1 = Never
  • 2 = Once
  • 3 = Twice
  • 4 = A few times
  • 5 = Very often

Level 2 Variables (L2)

Psychological Contract (Robinson & Morrison, 2000)

Indicate the extent to which you agree or disagree with the following statements regarding your experiences with your current employer.

Response Anchors: Agreement-based

  • 1 = Strongly disagree
  • 2 = Somewhat disagree
  • 3 = Neither agree nor disagree
  • 4 = Somewhat agree
  • 5 = Strongly agree
Expand for items

Perceived Contract Breach

  1. Almost none of the promises made by my employer have been kept so far.
  2. I feel that my employer has failed to fulfill the promises made to me when I was hired.
  3. So far, my employer has done a poor job of fulfilling its promises to me.
  4. I have not received everything my employer promised to me in exchange for my contributions.
  5. My employer has broken many of its promises to me, even though I have upheld my side of the deal.

Violation

  1. I feel a significant amount of anger toward my organization.
  2. I feel deeply betrayed by my organization.
  3. I feel that my organization has seriously violated the contract between us.
  4. I feel extremely frustrated with how I have been treated by my organization.

Job Satisfaction (Fisher et al., 2016, p. 20)

Overall, I am satisfied with my job.

Response Anchors: Agreement-based

  • 1 = Strongly disagree
  • 2 = Somewhat disagree
  • 3 = Neither agree nor disagree
  • 4 = Somewhat agree
  • 5 = Strongly agree

International Positive and Negative Affect Schedule Short Form (I-PANAS-SF; Thompson, 2007)

Thinking about yourself and your general emotional tendencies, how often do you feel:

Response Anchors: Frequency-based

  • 1 = Never
  • 2 = Rather infrequently
  • 3 = Some of the time
  • 4 = Quite often
  • 5 = Always
Expand for items

Positive Affect

  1. Alert
  2. Inspired
  3. Determined
  4. Attentive
  5. Active

Negative Affect

  1. Upset
  2. Hostile
  3. Ashamed
  4. Nervous
  5. Afraid

Demographic Variables (Hughes et al., 2022)

Expand for items

  1. Age (What is your age?)

    • Integer

  2. Ethnicity (What best describes your ethnic background?)

    • American Indian or Alaska Native
    • Asian
    • Black or African American
    • Arab, Middle Eastern, or North African
    • Hispanic or Latino
    • Native Hawaiian or Other Pacific Islander
    • White or Caucasian
    • Two or more races
    • Prefer not to say

  3. Gender (What best describes your gender identity?)

    • Woman
    • Man
    • Non-binary
    • Other
    • Prefer not to say

  4. Job Tenure (How long have you been in your current role?)

    • Less than a year
    • 1 to 3 years
    • 3 to 5 years
    • More than 5 years

  5. Education Level (What is the highest degree or level of school you have completed?)**

    • Some high school, no diploma
    • High school diploma or the equivalent (e.g., GED)
    • Some college, no degree
    • Vocational training
    • Associate degree
    • Bachelor’s degree
    • Master’s degree
    • Professional or Doctorate degree

  6. Remote Flag (I will be working remotely on the date I selected to receive the follow-up surveys.)

    • True
    • False

Analytical Strategy (Proposed)

Descriptive Statistics

  • Missing data analysis (moot since only complete responses will be used)
  • Outlier detection and handling (e.g., z-scores, boxplots/violent plots, etc.)
  • Careless responding analysis (Meade & Craig, 2012)
  • Means, standard deviations, and ranges for all variables
  • Normality checks for continuous variables (e.g., skewness and kurtosis, Shapiro-Wilk test, etc.)
  • Frequencies for categorical variables
    • Visualizations of group differences (e.g., bar plots, violin plots, etc.)
    • Effect size calculations (e.g., Cohen's d, eta-squared) for group differences
  • Visualizations of data distributions (e.g., histograms, boxplots, etc.)
  • Test for homoscedasticity (i.e., group differences)
  • Correlations between variables and visualizations of relationships
    • Pearson's r for continuous variables (L1 variables will be aggregated up to L2 level)
    • rmcorr exclusively for continuous variables at the L1 level
  • Intraclass Correlation Coefficients (ICCs) to assess within- and between-person variability (this will also determine if the within-person variables show intraindividual variability)
    • Low ICC implies more within-person fluctuation, high ICC implies more between-person stability
    • If within-person predictor meaningfully varies, include person mean at Level 2
    • If within-person predictor does not vary, treat it as a Level 2 predictor instead (or drop the within-person term for that variable)
    • In order to test within- and between-person difference, grand mean center the Level 1 tim-varying variable and include the person mean at Level 2; this parameterization directly tests the contextual difference (BP - WP) and is algebraically equivalent to person-mean centering under fixed within-person slopes.

Reliability

  • Multilevel composite reliability indices estimated across all levels (total, between-person, and within-person) via multilevel confirmatory factor analysis (MCFA) using the lavaan and semTools R packages (Lai, 2021; Geldhof et al., 2014)

Common Method Variance & Factor Structure

  • Multilevel CFA (MCFA) to simultaneously verify the theoretical factor structure at Level 1 and Level 2
  • CFA Marker Technique to evaluate and control for common method bias (see Williams, Hartman, & Cavazotte, 2010)

A Priori Sensitivity Analysis of Power

The author developed an R program to run robust sensitivity analyses of power based on the work of Arend & Schäfer, 2019. The program leverages the furrr (for parallelization) and simr packages in R. The original work of Arend & Schäfer, 2019 was expanded to focus on a (customizable) matrix comprising the following parameters (values listed within parentheses are used in the current study):

Parameter Variations:

  • Level 1 sample size (3)
  • Level 2 sample sizes (100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1100, 1200, 1300, 1400, 1500)
  • Level 1 direct effect size (0.10, 0.30, 0.50)
  • Level 2 direct effect size (0.10, 0.30, 0.50)
  • Cross-level effect size (0.10, 0.30, 0.50)
  • ICC values (0.10, 0.30, 0.50)
  • Random slope variance values (0.01, 0.09, 0.25)
  • Significance level (0.05)

The current implementation uses a 15 × 3 × 3 × 3 × 3 × 3 = 3,645 design matrix where each cell represents a unique combination of parameters. The simulations are run for 1,000 iterations per cell, resulting in a total of 3,645,000 simulations. To review the code and get a detailed breakdown along with instructions on how to run the process, see the program’s README.

Each power analysis calculated detection probability for three effects:

  • L1 Direct Effect
  • L2 Direct Effect
  • Cross-Level Interaction

By systematically varying key assumptions (ICC, effect sizes, random slope variance, etc.) and running simulations, the program provides insights into the sensitivity of power estimates under different conditions. Based on the results, the desired Level 2 sample size for the current study is 800 participants, which achieves adequate power (> 90%) to detect practically meaningful effects at both levels when holding other components at medium values. Figure 1 illustrates the power curves generated from the simulations.

power curves

Mixed-Effects (Multilevel) Modeling

All hypotheses will be tested using multilevel modeling via the lme4 R package (Bates et al., 2015; see also McNeish & Matta, 2018). Level 1 predictors will be person-mean (group) centered; Level 2 predictors will be grand-mean centered. A sequential model-building approach is used:

  1. Null model (unconditional means) — establishes ICC baseline and tests whether grouping structure accounts for variance in TI
  2. H1a/H2a — add L1 within-person predictors (NF subdimensions, burnout subdimensions, meeting load)
  3. H1b/H2b — add L2 between-person means (NF/burnout person-means)
  4. H4a/H4b/H5 — add L2 study variables (PC breach, violation, job satisfaction) and demographic covariates
  5. H3a/H3b — add meeting × NF and meeting × burnout within-person interaction terms; burnout and NF composites are computed for construct-level moderation testing

Rather than conducting significance tests to evaluate slope variance (which are unreliable with only 3 timepoints), an effect size estimate called ICC beta (ρ_β) will be calculated to evaluate the proportion of within-group variance in TI attributable to between-group differences in slopes (Aguinis & Culpepper, 2015). ICC beta is orthogonal to the standard ICC and will be reported in place of formalized slope variance tests.

Covariates modeled as between-person controls: age, job tenure, positive affect, and negative affect.

Empty/Null/Unconditional Means, Random Intercept:

Note. Replace TI with all other Level 1 variables (i.e., BURN.PHY, NF.COMP, etc.) to get the respective ICCs.

$$ \begin{aligned} \text{Level 1:}& \quad TI_{ti} = \beta_{0i} + \epsilon_{ti} \\ \text{Level 2:}& \quad \beta_{0i} = \gamma_{00} + \mu_{0i} \end{aligned} $$

Composite Notation:

$$ \begin{aligned} \quad TI_{ti} = \gamma_{00} + \mu_{0i} + \epsilon_{ti} \end{aligned} $$

Random Intercepts, Fixed Slopes (Full Research Model):

The within-person predictors are person-mean centered and is indicated by "WP" in the notation (e.g., $X_{ti} - \overline X_{i}$). The between-person predictors are grand mean centered and is indicated by "BP" in the notation (e.g., $X_{i} - \overline X$). Inspiration drawn from Perplexity of *Du, C. et. al (2024). Note: $z_{i}$ is a placeholder for covariates but are consolidated for LaTex rendering purposes.

$$ \begin{aligned} \text{Level 1:}& \quad TI_{ti} = \beta_{0i} + \beta_{1i}(BURN.PHY.WP_{ti}) + \beta_{2i}(BURN.COG.WP_{ti}) + \beta_{3i}(BURN.EMO.WP_{ti}) + \beta_{4i}(NF.COMP.WP_{ti}) + \beta_{5i}(NF.AUTO.WP_{ti}) + \beta_{6i}(NF.RLTD.WP_{ti})\\ & \quad\quad\quad\ + \beta_{7i}(MEET.COUNT.WP_{ti}) + \beta_{8i}(MEET.TIME.WP_{ti}) \\ & \quad\quad\quad\ + \beta_{9i}(MEET.COUNT.WP_{ti} \times BURN.PHY.WP_{ti}) + \beta_{10i}(MEET.COUNT.WP_{ti} \times BURN.COG.WP_{ti}) + \beta_{11i}(MEET.COUNT.WP_{ti} \times BURN.EMO.WP_{ti}) \\ & \quad\quad\quad\ + \beta_{12i}(MEET.COUNT.WP_{ti} \times NF.COMP.WP_{ti}) + \beta_{13i}(MEET.COUNT.WP_{ti} \times NF.AUTO.WP_{ti}) + \beta_{14i}(MEET.COUNT.WP_{ti} \times NF.RLTD.WP_{ti}) \\ & \quad\quad\quad\ + \epsilon_{ti} \\ &\quad \text{where } \epsilon_{ti} \sim N(0, \sigma_{\epsilon}^2) \\ \\ \end{aligned} $$

$$ \begin{aligned} \text{Level 2:}& \quad \beta_{0i} = \gamma_{00} + \gamma_{01}(BURN.PHY.BP_{i}) + \gamma_{02}(BURN.COG.BP_{i}) + \gamma_{03}(BURN.EMO.BP_{i}) + \gamma_{04}(NF.COMP.BP_{i}) + \gamma_{05}(NF.AUTO.BP_{i}) + \gamma_{06}(NF.RLTD.BP_{i}) \\ & \quad\quad\quad\ + \gamma_{07}(MEET.COUNT.BP_{i}) + \gamma_{08}(MEET.TIME.BP_{i}) \\ & \quad\quad\quad\ + \gamma_{09}(PSYK.BR_{i}) + \gamma_{10}(PSYK.VIO_{i}) + \gamma_{11}(JOBSAT_{i}) + \gamma_{12}(z_{i}) \\ & \quad\quad\quad\ + \mu_{0i} \\ &\quad \text{where } \mu_{0i} \sim N(0, \sigma_{\mu_{0}}^2) \\ \\ & \quad \beta_{1i} = \gamma_{10} \\ & \quad \beta_{2i} = \gamma_{20} \\ & \quad \beta_{3i} = \gamma_{30} \\ & \quad \beta_{4i} = \gamma_{40} \\ & \quad \beta_{5i} = \gamma_{50} \\ & \quad \beta_{6i} = \gamma_{60} \\ & \quad \beta_{7i} = \gamma_{70} \\ & \quad \beta_{8i} = \gamma_{80} \\ & \quad \beta_{9i} = \gamma_{90} \\ & \quad \beta_{10i} = \gamma_{100} \\ & \quad \beta_{11i} = \gamma_{110} \\ & \quad \beta_{12i} = \gamma_{120} \\ & \quad \beta_{13i} = \gamma_{130} \\ & \quad \beta_{14i} = \gamma_{140} \end{aligned} $$

Composite Notation:

$$ \begin{aligned} TI_{ti} =; & \gamma_{00} + \gamma_{01}(BURN.PHY.BP_{i}) + \gamma_{02}(BURN.COG.BP_{i}) + \gamma_{03}(BURN.EMO.BP_{i}) + \gamma_{04}(NF.COMP.BP_{i}) + \gamma_{05}(NF.AUTO.BP_{i}) + \gamma_{06}(NF.RLTD.BP_{i}) \\ & + \gamma_{07}(MEET.COUNT.BP_{i}) + \gamma_{08}(MEET.TIME.BP_{i}) \\ & + \gamma_{09}(PSYK.BR_{i}) + \gamma_{10}(PSYK.VIO_{i}) + \gamma_{11}(JOBSAT_{i}) + \gamma_{12}(z_{i}) \\ & + \gamma_{10}(BURN.PHY.WP_{ti}) + \gamma_{20}(BURN.COG.WP_{ti}) + \gamma_{30}(BURN.EMO.WP_{ti}) \\ & + \gamma_{40}(NF.COMP.WP_{ti}) + \gamma_{50}(NF.AUTO.WP_{ti}) + \gamma_{60}(NF.RLTD.WP_{ti}) \\ & + \gamma_{70}(MEET.COUNT.WP_{ti}) + \gamma_{80}(MEET.TIME.WP_{ti}) \\ & + \gamma_{90}(MEET.COUNT.WP_{ti} \times BURN.PHY.WP_{ti}) + \gamma_{100}(MEET.COUNT.WP_{ti} \times BURN.COG.WP_{ti}) + \gamma_{110}(MEET.COUNT.WP_{ti} \times BURN.EMO.WP_{ti}) \\ & + \gamma_{120}(MEET.COUNT.WP_{ti} \times NF.COMP.WP_{ti}) + \gamma_{130}(MEET.COUNT.WP_{ti} \times NF.AUTO.WP_{ti}) + \gamma_{140}(MEET.COUNT.WP_{ti} \times NF.RLTD.WP_{ti}) \\ & + \mu_{0i} + \epsilon_{ti} \end{aligned} $$

Limitations

Primarily, vague response anchors were used to assess frequency. While Tong et al. (2020) demonstrated that precise response options yield superior performance, the expedited timeline of this study necessitated the use of less specific anchors. In future research, particularly longitudinal studies spanning months or years or cross-sectional designs, employing more precise response options would likely enhance measurement reliability and validity.

The outcome variable, turnover intentions, was measured using a single item. While this approach is justified in the context of this study (see Matthews et al., 2022), it is important to acknowledge that single-item measures may not capture the full complexity of the construct. Future research should consider using multi-item measures to provide a more comprehensive assessment of turnover intentions. In addition, turnover intentions and all other variables used in the study were measured using a 5-point agreement-based Likert scale (1 to 5), which are ordinal in nature. While this approach is common in psychological research, it is important to acknowledge that treating ordinal data as interval data can lead to potential biases in parameter estimates and attenuated effect sizes.

All variables used in the study were self-reported, which raises concerns about common method variance (CMV) or same-source bias. This limitation is particularly relevant in the context of psychological constructs, where self-report measures are often used. While the study employed statistical techniques to assess and mitigate the impact of CMV, it is important to recognize that these methods cannot completely eliminate the potential influence of same-source bias on the findings (see Williams, Hartman, & Cavazotte, 2010). Future research should consider using multi-source data collection or longitudinal designs with lagged outcomes to address this concern more effectively. Although statistical analyses were conducted to assess the extent of same-source bias or common method bias, these methods cannot entirely eliminate its potential influence on the findings (See Williams, Hartman, & Cavazotte, 2010*). Because this study was not an extensive longitudinal design, I decided not to use lagged outcomes as controls. It is plausible that common method variance may have influenced the results given the concurrent relationships between the predictors and turnover intentions (Gabriel et. al, 2019, Kim et. al, 2025). Future research should consider employing a more robust design to mitigate this concern, such as using multi-source data collection or longitudinal designs with lagged outcomes.

The sample is likely WEIRD (Western, Educated, Industrialized, Rich, Democratic) and limited to participants working standard 9-to-5 shifts. It is plausible that differences may emerge across non-standard shifts, warranting further investigation to explore potential variations (See Alley & Cortina, 2023).

The power analysis was conducted using a custom R program that simulates data based on the proposed model but including a single L1 and L2 variable. While this approach provides a robust framework for estimating power, it is important to note that the results are contingent upon the assumptions made in the simulation; the inclusion of multiple predictors typically means an increase in multicollinearity which leads to more inflated standard error estimates (i.e., reduced power). Arend & Schäfer (2019) state that multicollinearity has to be very high in order to cause a noticeable downward bias in power estimates (i.e., VIF ≥ 2 and VIF ≥ 5), so VIFs are reported to address this concern. Variations in effect sizes, ICC values, and other parameters can also significantly influence the power estimates, albeit some are more influential than others. Therefore, while the analysis offers valuable insights, it should be interpreted with caution and in the context of the specific research design and hypotheses.

Latent mean centering is recognized as a superior technique for disentangling fixed and random effects. However, implementing this approach in R presented significant challenges. A Bayesian framework may offer a more robust alternative by modeling means as distributions, thereby directly accounting for uncertainty in the estimates (see Hamaker & Muthén, 2020). In addition, sampling error in the group means exists because the literal group means were used as the L2 estimates. As with the previous point, there are ways to address this (empirical Bayes) but none are easily implemented with the R software..hence the limitation. See this article by Curran and Bauer.

The specified model(s) in the study are based on fixed slopes; sampling and budgetary restrictions prevented me from conducting a more intensive longitudinal study which led to the decision to only model fixed slopes. Additionally, check out this link which explains why one should check for random slopes in the predictors as well Perplexity.

Need satisfaction and need frustration are known to predict positive and negative outcomes differentially. This study focuses specifically on need thwarting or frustration, thereby capturing only one aspect of the broader dynamic. Future research should aim to provide a more balanced perspective by examining both need satisfaction and frustration. Perhaps a configurational (also known as person-oriented) approach could be employed to explore the interplay between need satisfaction, frustration, and the other relvenat study variables across time (i.e., multilevel finite mixture modeling).

At least three measurement occasions are necessary to estimate within-person effects. Stronger within-person effects might have been observed with more Level-1 data; perhaps there is a curvilinear (i.e., quadratic) trend in turnover intentions across the measurement occasions for which the present study did not account, but with so few measurement occasions, it is unlikely that a quadratic trend would be detected. Future research should consider including more measurement occasions to explore potential curvilinear relationships.

The study design could not directly assess the causal direction between the predictors and turnover intentions. While the proposed model suggests a directional relationship, it is important to acknowledge that the data are correlational in nature. Future research should consider employing experimental or quasi-experimental designs to establish causal relationships more definitively.

The sample size is too small to test the moderating relationships in a single model; in fact, some of the global constructs might have to be used due to power issues. Separate models will most likely have to be tested to get an understanding of the phenomena.

About

Doctoral thesis research as DataOps & DevOps.

Resources

Readme
License
Activity

Stars

1 star

Watchers

1 watching

Forks

0 forks
Report repository

Releases

Packages

Used by

Contributors

Languages