diff --git a/CHANGELOG.md b/CHANGELOG.md
index 044d667..5890f01 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -20,6 +20,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Added Features and Improvements 🙌:
 - Supporting latest Python 3.11 🎉
 
+### Bugfix 🐛:
+- Fixed wrong connectivity in `mh.datasets.utils.propagate_MCMC` for different datatypes, #39
+
 
 ## [1.0.4] - 2023-05-22
 ### Other changes:
diff --git a/src/msmhelper/msm/timescales.py b/src/msmhelper/msm/timescales.py
index dc3ecef..c0151c6 100644
--- a/src/msmhelper/msm/timescales.py
+++ b/src/msmhelper/msm/timescales.py
@@ -321,15 +321,15 @@ def _propagate_MCMC_step(cummat, idx_from):
     """Propagate a single step Markov chain Monte Carlo."""
     rand = random.random()  # noqa: S311
     cummat_perm, state_perm = cummat
-    cummat_perm, state_perm = cummat_perm[idx_from], state_perm[idx_from]
+    cummat_from, state_from = cummat_perm[idx_from], state_perm[idx_from]
 
-    for idx, cummat_idx in enumerate(cummat_perm):
+    for idx, cummat_idx in enumerate(cummat_from):
         # strict less to ensure that rand=0 does not jump along unconnected
         # states with Tij=0.
         if rand < cummat_idx:
-            return state_perm[idx]
+            return state_from[idx]
     # this should never be reached, but needed for numba to ensure int return
-    return len(cummat_perm) - 1
+    return state_from[np.argmax(cummat_from)]  # pragma: no cover
 
 
 @numba.njit
diff --git a/src/msmhelper/utils/datasets.py b/src/msmhelper/utils/datasets.py
index 1a8d5de..e92cc2e 100644
--- a/src/msmhelper/utils/datasets.py
+++ b/src/msmhelper/utils/datasets.py
@@ -7,6 +7,7 @@
 
 import numpy as np
 
+from msmhelper.msm.msm import row_normalize_matrix
 from msmhelper.msm.timescales import _propagate_MCMC
 from .tests import is_transition_matrix
 from ._utils import shift_data
@@ -47,8 +48,10 @@ def propagate_tmat(tmat, nsteps, start=None):
         raise ValueError('tmat needs to be a row-normalized matrix.')
 
     n_states = len(tmat)
-    cummat = np.cumsum(tmat, axis=1)
-    cummat[:, -1] = 1  # enforce exact normalization
+    cummat = np.cumsum(  # enforce exact normalization
+        row_normalize_matrix(tmat),
+        axis=1,
+    )
     cummat_perm = np.tile(np.arange(n_states), (n_states, 1))
 
     if start is None:
diff --git a/test/test_utils_datasets.py b/test/test_utils_datasets.py
new file mode 100644
index 0000000..fb47045
--- /dev/null
+++ b/test/test_utils_datasets.py
@@ -0,0 +1,50 @@
+# -*- coding: utf-8 -*-
+"""Tests for the filtering submodule.
+
+BSD 3-Clause License
+Copyright (c) 2019-2023, Daniel Nagel
+All rights reserved.
+
+"""
+import numpy as np
+import pytest
+from msmhelper import msm
+from msmhelper.utils import datasets
+
+NSTEPS = 1000000
+DECIMAL = 2
+
+
+def test_nagel20_4state():
+    """Test nagel20_4state mcmc."""
+    traj_mcmc = datasets.nagel20_4state(NSTEPS)
+
+    np.testing.assert_array_almost_equal(
+        datasets.nagel20_4state.tmat,
+        msm.estimate_markov_model(trajs=traj_mcmc, lagtime=1)[0],
+        decimal=DECIMAL,
+    )
+
+
+def test_nagel20_6state():
+    """Test nagel20_6state mcmc."""
+    traj_mcmc = datasets.nagel20_6state(NSTEPS)
+
+    np.testing.assert_array_almost_equal(
+        datasets.nagel20_6state.tmat,
+        msm.estimate_markov_model(trajs=traj_mcmc, lagtime=1)[0],
+        decimal=DECIMAL,
+    )
+
+
+@pytest.mark.parametrize('nstates', [2, 2, 3, 4, 5])
+def test_propagate_tmat(nstates):
+    tmat = np.random.uniform(size=(nstates, nstates))
+    tmat = msm.row_normalize_matrix(tmat)
+
+    traj_mcmc = datasets.propagate_tmat(tmat, NSTEPS)
+    np.testing.assert_array_almost_equal(
+        tmat,
+        msm.estimate_markov_model(trajs=traj_mcmc, lagtime=1)[0],
+        decimal=DECIMAL,
+    )